Enhancing healthcare classification with hybrid multimedia data processing and deep learning TNBO FCNN approach in IoT-enabled environments Authors: Leeladhar Chourasiya, Umesh Kumar Lillohre, Abhishek Kumar Pandey, Sarita Simaiya, Shimaa A. Hussien, Ehab Seif Ghith, Monish Khan Published: 2025 ABSTRACT The rapid expansion of IoT-based healthcare systems has led to an explosion of multimodal patient data (e.g., voice, images, signals), raising concerns about accurate classification, real-time processing, and data security. Existing machine learning approaches often suffer from limited feature extraction, high computational demands, and inadequate privacy protection, particularly in decentralised and resource-constrained environments. Furthermore, the lack of robust routing and data integrity mechanisms hinders their deployment in real-world healthcare applications. To address these challenges, we propose a hybrid deep learning framework that integrates Tunable Nonlinear Bayesian Optimisation (TNBO) for efficient routing and hyper-parameter tuning with a Fully Connected Neural Network (FCNN) for robust classification of multimodal healthcare data. Additionally, we embed blockchain technology to ensure the security, transparency, and immutability of patient data across decentralised IoT environments. Our method demonstrates superior performance in terms of accuracy, sensitivity, and specificity when compared to traditional models such as RTS-DELM and SCA_WKNN. We optimise the proposed architecture for energy efficiency, scalability, and real-time healthcare monitoring, positioning it for deployment in telemedicine and smart healthcare systems. Experimental validation on benchmark datasets confirms the efficacy of our model, while ethical implications and future research directions are also discussed to support broader applicability. The performance of the proposed model was evaluated against existing state-of-the-art techniques, including Hybrid Multimedia Data Processing, RTS-DELM, Multimedia Data Sharing Systems, and SCA_WKNN. The results indicated that the TNBO + FCNN model achieved an accuracy of 0.924, a sensitivity of 0.921, and a specificity of 0.926. Furthermore, the model outperformed comparative approaches by significant margins, achieving an F1-score of 0.915 and an area under the receiver operating characteristic curve (AUC-ROC) of 0.950. These findings validate the proposed model’s effectiveness in enhancing healthcare data classification, making it a promising solution for real-time monitoring and decision-making in healthcare systems.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-33141-0. Introduction The Internet of Things (IoT) has significantly impacted sectors such as agriculture, smart homes, and especially healthcare by enabling devices to seamlessly connect and exchange data1,2. In healthcare, IoT supports remote monitoring, efficient diagnostics, and treatment tracking through a network of embedded, interconnected devices3,4. Applications like smart grids and IoT-based patient monitoring systems are transforming healthcare delivery and addressing global health disparities5. Continuous IoT advancements have enabled real-time data transmission, boosted diagnostic accuracy and improved patient outcomes6,7. Blockchain technology further enhances this ecosystem by securing electronic health records and enabling privacy-preserving, tamper-proof data exchange8,9. Its immutable and decentralised nature strengthens trust in data-driven medical analytics and reduces the need for excessive medical interventions in chronic illness management10. In parallel, multimodal healthcare systems which integrate images, audio, tabular data, and time series support complex decision-making by simulating clinical reasoning. Such approaches have shown promise across various domains, including radiology, cardiology, psychiatry, and public health8. As healthcare becomes more digitized and decentralized, risks to data privacy rise due to software vulnerabilities, insider threats, and insecure smart devices11. Blockchain addresses these challenges by using cryptographic hashes to ensure data integrity. Several recent models, such as RTS-DELM12 and SCA-based Weighted K-NN13,15, demonstrate the evolution of intelligent classification and prediction systems. However, limitations remain in handling multimodal inputs and securing data effectively. To address this, our research proposes an IoT-blockchain-based system comprising device, communication, blockchain, and user interface layers. In the computation layer, spiral handwriting images and voice signals undergo processing, while the TNBO algorithm guarantees energy-efficient routing. The proposed FCNN model16 follows preprocessing, data augmentation, and feature extraction for classification. A significant contribution is illustrated as follows: Optimised Routing for Blockchain-Enabled IoT Using the Proposed TNBO: The proposed TNBO is utilised in the routing process of blockchain-enabled IoT. Here, the Tangent Search Algorithm (TSA) and Namib Beetle Optimisation (NBO) were combined to create the proposed TNBO.Healthcare Classification Using the Proposed FCNN Model: The hybrid FCNN model is designed for IoT-blockchain-based healthcare classification. Incorporating a convolutional neural network with fused features creates the proposed FCNN model. Optimised Routing for Blockchain-Enabled IoT Using the Proposed TNBO: The proposed TNBO is utilised in the routing process of blockchain-enabled IoT. Here, the Tangent Search Algorithm (TSA) and Namib Beetle Optimisation (NBO) were combined to create the proposed TNBO. Optimised Routing for Blockchain-Enabled IoT Using the Proposed TNBO: The proposed TNBO is utilised in the routing process of blockchain-enabled IoT. Here, the Tangent Search Algorithm (TSA) and Namib Beetle Optimisation (NBO) were combined to create the proposed TNBO. Healthcare Classification Using the Proposed FCNN Model: The hybrid FCNN model is designed for IoT-blockchain-based healthcare classification. Incorporating a convolutional neural network with fused features creates the proposed FCNN model. Healthcare Classification Using the Proposed FCNN Model: The hybrid FCNN model is designed for IoT-blockchain-based healthcare classification. Incorporating a convolutional neural network with fused features creates the proposed FCNN model. The complete article is organised into various subsections, which primarily cover a literature review related to blockchain and deep learning-based healthcare research. Section three covers materials and methods, section four presents simulation results and discussion, and section five concludes with a discussion of the future direction of the work. Challenges in existing methods Despite the widespread adoption of machine learning and deep learning in healthcare data classification, several limitations persist. Conventional models often lack effective feature extraction for complex multimodal inputs (e.g., audio, images), and manual feature selection hampers performance17. High computational demands and the need for large training datasets restrict real-time applicability. Many healthcare datasets suffer from class imbalance, leading to biased predictions and reduced sensitivity toward minority classes. Furthermore, inadequate privacy protections and centralised data storage expose patient data to security risks. Most existing models also struggle to generalise across datasets and fail to integrate multiple data modalities effectively, resulting in inconsistent classification performance4,11. Problem statement and objectives Current healthcare classification methodologies face numerous challenges, including ineffective feature extraction, significant computational demands, data imbalance, and privacy concerns. Conventional machine learning and deep learning techniques often fail to efficiently handle multimodal data, which includes audio and image-based medical records2,18. Furthermore, the security and data integrity of centralised storage systems undermine the dependability of healthcare predictions. To mitigate these challenges, there is a pressing need for a robust and efficient classification model that can enhance accuracy, safeguard data, and process various types of medical data seamlessly. This research aims to develop an advanced healthcare classification model based on TNBO + FCNN, which enhances classification performance while addressing existing challenges19. The specific objectives are as follows: Enhancing Classification Accuracy: Develop a hybrid deep learning framework (TNBO + FCNN) to optimise feature extraction and classification processes. Utilise iterative optimisation techniques to improve the precision of healthcare predictions20.Creating an efficient training mechanism to reduce computational overhead. Ensure the model can handle large-scale healthcare datasets without excessive memory usage.Employ data augmentation techniques to balance the dataset and improve generalisation. Implement adaptive learning strategies to enhance sensitivity towards minority classes.Integrating Multimodal Learning Effectively combines audio and image-based medical data through a multimodal approach. Assess the contribution of various modalities to overall classification performance.Ensuring Data Security and Integrity. Incorporate blockchain technology to secure medical records and prevent unauthorised data alterations. Strengthen privacy-preserving mechanisms to ensure compliance with healthcare security regulations21.Using appropriate metrics, compare the proposed method to existing models (Hybrid Multimedia Processing, RTS-DELM, and SCA_WKNN). Enhancing Classification Accuracy: Develop a hybrid deep learning framework (TNBO + FCNN) to optimise feature extraction and classification processes. Utilise iterative optimisation techniques to improve the precision of healthcare predictions20. Enhancing Classification Accuracy: Develop a hybrid deep learning framework (TNBO + FCNN) to optimise feature extraction and classification processes. Utilise iterative optimisation techniques to improve the precision of healthcare predictions20. Creating an efficient training mechanism to reduce computational overhead. Ensure the model can handle large-scale healthcare datasets without excessive memory usage. Creating an efficient training mechanism to reduce computational overhead. Ensure the model can handle large-scale healthcare datasets without excessive memory usage. Employ data augmentation techniques to balance the dataset and improve generalisation. Implement adaptive learning strategies to enhance sensitivity towards minority classes. Employ data augmentation techniques to balance the dataset and improve generalisation. Implement adaptive learning strategies to enhance sensitivity towards minority classes. Integrating Multimodal Learning Effectively combines audio and image-based medical data through a multimodal approach. Assess the contribution of various modalities to overall classification performance. Integrating Multimodal Learning Effectively combines audio and image-based medical data through a multimodal approach. Assess the contribution of various modalities to overall classification performance. Ensuring Data Security and Integrity. Incorporate blockchain technology to secure medical records and prevent unauthorised data alterations. Strengthen privacy-preserving mechanisms to ensure compliance with healthcare security regulations21. Ensuring Data Security and Integrity. Incorporate blockchain technology to secure medical records and prevent unauthorised data alterations. Strengthen privacy-preserving mechanisms to ensure compliance with healthcare security regulations21. Using appropriate metrics, compare the proposed method to existing models (Hybrid Multimedia Processing, RTS-DELM, and SCA_WKNN). Using appropriate metrics, compare the proposed method to existing models (Hybrid Multimedia Processing, RTS-DELM, and SCA_WKNN). Motivation for the research The rapid advancement in AI and deep learning has significantly transformed healthcare applications, particularly in disease classification and diagnosis. Nonetheless, current classification models face numerous challenges, including ineffective feature extraction, high computational demands, concerns regarding data security, and difficulties in managing multimodal medical data, which encompasses audio and image records22. These challenges hinder the timely and accurate prediction of diseases, underscoring the need for a more robust, efficient, and secure classification model. This study is driven by several critical factors: the demand for Enhanced Classification Accuracy. Current models demonstrate limited accuracy due to subpar feature extraction from healthcare data. A deep learning approach, such as TNBO + FCNN, can optimise feature representation and improve classification outcomes. Challenges in Managing Multimodal Healthcare Data. Healthcare data encompasses a variety of formats, including medical images, audio diagnostics, and textual records23. Existing models frequently handle these modalities in isolation, resulting in less effective decision-making. A multimodal learning framework is essential for integrating diverse data types to improve classification accuracy. Addressing Computational Efficiency and Scalability: Numerous deep learning models require substantial computational resources, which limit their real-time applicability in healthcare settings. This research aims to develop a computationally efficient hybrid model that can process extensive datasets without incurring excessive resource expenditure. Ensuring Data Security and Integrity through Blockchain Technology. Medical records contain sensitive patient information, making security a paramount concern3,7. Incorporating blockchain technology can offer tamper-resistant storage solutions and bolster the reliability of healthcare data. Connecting Research with Practical Applications. Many AI-based healthcare solutions remain limited to academic environments due to scalability and practical implementation challenges. This research aims to bridge this gap. Key contributions In response to limitations in existing healthcare classification models, this study proposes a hybrid deep learning framework combining Tunable Nonlinear Bio-Inspired Optimisation (TNBO) with Fully Connected Neural Networks (FCNN). The contributions are as follows: A novel TNBO + FCNN architecture that improves feature selection and classification accuracy,A multimodal classification strategy that combines medical sound and imaging data for robust diagnostics, blockchain-based data security ensuring integrity, privacy, and secure data sharing, Superior performance over existing methods (e.g., RTS-DELM, SCA_WKNN) in accuracy, sensitivity, specificity, F1-score, and AUC-ROC, enhanced computational efficiency via TNBO, enabling faster training and real-time applicability.Comprehensive evaluation with varying training set sizes (60–95%) and K-fold cross-validation (K = 5, 7, 9, 11), validated against benchmark models for credible performance assessment. A novel TNBO + FCNN architecture that improves feature selection and classification accuracy, A novel TNBO + FCNN architecture that improves feature selection and classification accuracy, A multimodal classification strategy that combines medical sound and imaging data for robust diagnostics, blockchain-based data security ensuring integrity, privacy, and secure data sharing, Superior performance over existing methods (e.g., RTS-DELM, SCA_WKNN) in accuracy, sensitivity, specificity, F1-score, and AUC-ROC, enhanced computational efficiency via TNBO, enabling faster training and real-time applicability. A multimodal classification strategy that combines medical sound and imaging data for robust diagnostics, blockchain-based data security ensuring integrity, privacy, and secure data sharing, Superior performance over existing methods (e.g., RTS-DELM, SCA_WKNN) in accuracy, sensitivity, specificity, F1-score, and AUC-ROC, enhanced computational efficiency via TNBO, enabling faster training and real-time applicability. Comprehensive evaluation with varying training set sizes (60–95%) and K-fold cross-validation (K = 5, 7, 9, 11), validated against benchmark models for credible performance assessment. Comprehensive evaluation with varying training set sizes (60–95%) and K-fold cross-validation (K = 5, 7, 9, 11), validated against benchmark models for credible performance assessment. However, while the integration of TNBO with FCNN for multimodal healthcare data classification presents a novel contribution, the distinction from existing optimisation-based deep learning models requires clearer articulation to better establish its uniqueness and added value. Deep learning has shown promise in healthcare classification, but conventional models often rely on single-modality data, limiting diagnostic accuracy. This research introduces a multimodal approach that combines audio inputs (e.g., symptom recordings) with medical imaging (e.g., X-rays, MRIs) to provide a more comprehensive diagnostic view. The proposed TNBO + FCNN model effectively integrates these diverse data types, leveraging multimodal learning to enhance feature extraction and decision-making. This results in a more robust and reliable classification system capable of handling complex healthcare data. Literature review The integration of IoT and blockchain in healthcare enhances patient-provider interaction, enables real-time monitoring, and supports cost-effective treatment delivery. This study proposes a hybrid deep learning (DL) model for classifying healthcare data, emphasising secure, efficient, and scalable systems. Blockchain technology is recognised for ensuring confidentiality, integrity, and real-time control in healthcare environments. Taloba et al.1 introduced a blockchain-based security architecture using hashing to detect data breaches in multimedia health content. Rehman et al.2 combined blockchain with RTS-DELM and IDS to create a health monitoring framework for healthcare 5.0, although it lacked robustness against suboptimal data input. Similarly, Xie et al.3 demonstrated AI-enhanced medical data categorisation supported by blockchain for privacy and decentralised data exchange. Xi et al.4 and Tanwar et al.5 highlighted blockchain’s tamper-proof and traceable nature, making it suitable for secure EMR and EHR management. Sharma et al.6 extended this by integrating blockchain with IoMT to enhance medical data security and access. De Aguiar et al.7 emphasised blockchain’s role in interoperability and secure information exchange, while Goel and Neduncheliyan8 proposed DBDH—a deep learning-based blockchain security framework. Recent advancements in IoT-enhanced healthcare systems notably include frameworks for personalised medication management, such as the MediServe project, which leverages deep learning for elderly care (Kapse et al., 2024)32. Similarly, AI of Things applications have shown promise in the early detection of brain tumours via MRI imaging, further exemplifying multimodal diagnostic capabilities (Khekare et al., 2025)33. Moreover, optimising diagnostic accuracy for chest X-rays by fusing hyper-parameter tuning with transfer learning approaches has led to more reliable disease classification systems (Wajgi et al., 2024)34. The integration of smart IoT-enabled diagnostic tools continues to evolve, promoting better healthcare outcomes (Yenurkar et al., 2024a)35, while IoT-based blood pressure monitoring systems are transforming chronic heart disease management practices (Yenurkar et al., 2024b)36. Despite progress, the literature reveals key gaps: Hybrid DL models specifically tailored for IoT blockchain healthcare data classification remain underexplored1–8.Applications in chronic disease management, real-time monitoring, and personalised healthcare lack contextual depth2,3,5.There is insufficient reporting of system-level performance metrics like latency, throughput, and energy efficiency4,6.Limited studies address scalability and reliability, particularly under high data loads from IoT devices5,7.Advanced privacy-preserving frameworks and encryption methods are rarely evaluated in depth1,8. Hybrid DL models specifically tailored for IoT blockchain healthcare data classification remain underexplored1–8. Hybrid DL models specifically tailored for IoT blockchain healthcare data classification remain underexplored1–8. Applications in chronic disease management, real-time monitoring, and personalised healthcare lack contextual depth2,3,5. Applications in chronic disease management, real-time monitoring, and personalised healthcare lack contextual depth2,3,5. There is insufficient reporting of system-level performance metrics like latency, throughput, and energy efficiency4,6. There is insufficient reporting of system-level performance metrics like latency, throughput, and energy efficiency4,6. Limited studies address scalability and reliability, particularly under high data loads from IoT devices5,7. Limited studies address scalability and reliability, particularly under high data loads from IoT devices5,7. Advanced privacy-preserving frameworks and encryption methods are rarely evaluated in depth1,8. Advanced privacy-preserving frameworks and encryption methods are rarely evaluated in depth1,8. These gaps underscore the need for integrated, intelligent, and secure healthcare systems leveraging the full potential of IoT, blockchain, and deep learning. Challenges The challenges faced by existing techniques in the integration of IoT and blockchain technologies for healthcare data classification can be summarised as follows: Hybrid Multimedia Data Processing: Prioritising Patient Big Data. Implementing a hybrid multimedia data processing framework in telemedicine was complicated due to the need to prioritise and manage extensive patient data effectively. Additionally, constructing a medical management system for chronic diseases using blockchain technology posed significant challenges.RTS-DELM Approach: Difficulty in Factor Identification: The Real-Time Deep Extreme Learning System (RTS-DELM) approach demonstrated effectiveness but struggled to identify and quantify relevant factors influencing health outcomes. Moreover, the challenge of evaluating additional datasets added complexity to its implementation.Multimedia Data-Sharing Systems: Challenges in Implementing Blockchain. Despite the potential of multimedia data-sharing systems to create a secure, distributed detection model using blockchain technology, significant hurdles exist in implementing blockchain-based threat detection within machine learning (ML) algorithms. This limitation hindered the efficacy of threat detection mechanisms in healthcare settings.Sine Cosine Weighted K-Nearest Neighbour (SCA_WKNN): Limited Chronic Disease Detection: The SCA_WKNN approach addressed many Quality of Service (QoS) needs within healthcare systems. However, it fell short of effectively detecting chronic diseases that affect internal body parts. This limitation highlights the need for more effective techniques to address a broader range of health conditions.Obstacles in Medical Data Transmission: Challenges in Wireless IoT Networks: The transmission of multimedia data among wireless IoT networks faced numerous obstacles, primarily due to the reliance on centralised systems. These centralised frameworks not only posed significant security risks but also restricted user access to data, underscoring the need for more decentralised and secure systems. Hybrid Multimedia Data Processing: Prioritising Patient Big Data. Implementing a hybrid multimedia data processing framework in telemedicine was complicated due to the need to prioritise and manage extensive patient data effectively. Additionally, constructing a medical management system for chronic diseases using blockchain technology posed significant challenges. Hybrid Multimedia Data Processing: Prioritising Patient Big Data. Implementing a hybrid multimedia data processing framework in telemedicine was complicated due to the need to prioritise and manage extensive patient data effectively. Additionally, constructing a medical management system for chronic diseases using blockchain technology posed significant challenges. RTS-DELM Approach: Difficulty in Factor Identification: The Real-Time Deep Extreme Learning System (RTS-DELM) approach demonstrated effectiveness but struggled to identify and quantify relevant factors influencing health outcomes. Moreover, the challenge of evaluating additional datasets added complexity to its implementation. RTS-DELM Approach: Difficulty in Factor Identification: The Real-Time Deep Extreme Learning System (RTS-DELM) approach demonstrated effectiveness but struggled to identify and quantify relevant factors influencing health outcomes. Moreover, the challenge of evaluating additional datasets added complexity to its implementation. Multimedia Data-Sharing Systems: Challenges in Implementing Blockchain. Despite the potential of multimedia data-sharing systems to create a secure, distributed detection model using blockchain technology, significant hurdles exist in implementing blockchain-based threat detection within machine learning (ML) algorithms. This limitation hindered the efficacy of threat detection mechanisms in healthcare settings. Multimedia Data-Sharing Systems: Challenges in Implementing Blockchain. Despite the potential of multimedia data-sharing systems to create a secure, distributed detection model using blockchain technology, significant hurdles exist in implementing blockchain-based threat detection within machine learning (ML) algorithms. This limitation hindered the efficacy of threat detection mechanisms in healthcare settings. Sine Cosine Weighted K-Nearest Neighbour (SCA_WKNN): Limited Chronic Disease Detection: The SCA_WKNN approach addressed many Quality of Service (QoS) needs within healthcare systems. However, it fell short of effectively detecting chronic diseases that affect internal body parts. This limitation highlights the need for more effective techniques to address a broader range of health conditions. Sine Cosine Weighted K-Nearest Neighbour (SCA_WKNN): Limited Chronic Disease Detection: The SCA_WKNN approach addressed many Quality of Service (QoS) needs within healthcare systems. However, it fell short of effectively detecting chronic diseases that affect internal body parts. This limitation highlights the need for more effective techniques to address a broader range of health conditions. Obstacles in Medical Data Transmission: Challenges in Wireless IoT Networks: The transmission of multimedia data among wireless IoT networks faced numerous obstacles, primarily due to the reliance on centralised systems. These centralised frameworks not only posed significant security risks but also restricted user access to data, underscoring the need for more decentralised and secure systems. Obstacles in Medical Data Transmission: Challenges in Wireless IoT Networks: The transmission of multimedia data among wireless IoT networks faced numerous obstacles, primarily due to the reliance on centralised systems. These centralised frameworks not only posed significant security risks but also restricted user access to data, underscoring the need for more decentralised and secure systems. These challenges underscore the complexities and limitations existing technologies face in the healthcare sector, particularly when integrating IoT and blockchain. Addressing these challenges is crucial for developing practical solutions that enhance data management, security, and healthcare outcomes. Table 1 presents the comparative analysis of existing healthcare classification models. Table 1Comparative analysis of existing healthcare classification Models.Authors/ReferenceTechniqueDataset/Domain% AccuracyNotes/LimitationsRehman et al.RTS-DELM + BlockchainHealthcare 5.0 IDS83.7–93.22High accuracy, robust for the IDS datasetLakshmanan et al.TNBO FCNN + BlockchainParkinson’s & mPower Voice92.4–93.4Robust for multimodal health dataAlmogadwy et al.RTS-DELM (FFL fusion)Chronic Kidney Disease98.21Chronic kidney disease predictionMohamed et al.SCAWKNN/MCODBC-HDDCHD datasetSCAWKNN: 95.30, MCODBC-HDDC: 97.36Hybrid model, detailed comparative ablationLal et al.Hybrid Linguistic ML (SbDNM)Healthcare NLP & image96.8Hybrid with an optimised deep neural modelRaj et al.Big Data/IoT-based MLSensor fusion97.0IoT health monitoring, validated via sensor dataRatta et al.Blockchain + MLCOVID-19 monitoring (IoT)95.0COVID-19 diagnosis, IoT-based, robust securityTobón et al.Deep Learning MultimediaAudio/Image/Voice multimodality85.0–91.6Multimodal, robust for image/audio fusionSrinivasan et al.Secure Multimedia DataIoT medical records100.0Feed-forward neural network appliedLakshmanan et al.Fused FCNN (Blockchain-IoT)Healthcare multimodal92.4–93.4Comparable with other DL studies, multimodalProposed TNBO + FCNNTNBO + FCNN with Blockchain, Multimodal LearningParkinson’s Drawings + mPower Voice datasetsAchieved superior classification performance across all metrics (F1, AUC-ROC, sensitivity, etc.)Moderate computational demands; needs further testing on diverse conditions. Table 1Comparative analysis of existing healthcare classification Models.Authors/ReferenceTechniqueDataset/Domain% AccuracyNotes/LimitationsRehman et al.RTS-DELM + BlockchainHealthcare 5.0 IDS83.7–93.22High accuracy, robust for the IDS datasetLakshmanan et al.TNBO FCNN + BlockchainParkinson’s & mPower Voice92.4–93.4Robust for multimodal health dataAlmogadwy et al.RTS-DELM (FFL fusion)Chronic Kidney Disease98.21Chronic kidney disease predictionMohamed et al.SCAWKNN/MCODBC-HDDCHD datasetSCAWKNN: 95.30, MCODBC-HDDC: 97.36Hybrid model, detailed comparative ablationLal et al.Hybrid Linguistic ML (SbDNM)Healthcare NLP & image96.8Hybrid with an optimised deep neural modelRaj et al.Big Data/IoT-based MLSensor fusion97.0IoT health monitoring, validated via sensor dataRatta et al.Blockchain + MLCOVID-19 monitoring (IoT)95.0COVID-19 diagnosis, IoT-based, robust securityTobón et al.Deep Learning MultimediaAudio/Image/Voice multimodality85.0–91.6Multimodal, robust for image/audio fusionSrinivasan et al.Secure Multimedia DataIoT medical records100.0Feed-forward neural network appliedLakshmanan et al.Fused FCNN (Blockchain-IoT)Healthcare multimodal92.4–93.4Comparable with other DL studies, multimodalProposed TNBO + FCNNTNBO + FCNN with Blockchain, Multimodal LearningParkinson’s Drawings + mPower Voice datasetsAchieved superior classification performance across all metrics (F1, AUC-ROC, sensitivity, etc.)Moderate computational demands; needs further testing on diverse conditions. Comparative analysis of existing healthcare classification Models. Table 1 presents a comparative summary of recent healthcare classification models, focusing only on those references that report explicit “% Accuracy” values. The table lists several methods, datasets, accuracy rates, and notes about their strengths or limitations, enabling clear benchmarking of various deep learning and hybrid techniques used in IoT-enabled healthcare systems. Methodology The selection of TNBO for hyper-parameter optimisation and FCNN for handling complex, high-dimensional healthcare data forms a robust framework for multimedia health data classification. Their combined use offers superior computational efficiency and accuracy compared to traditional models. Additionally, integrating blockchain technology into this framework enhances data integrity and security by preventing unauthorised modifications. While blockchain does not directly impact classification accuracy, it ensures the trustworthiness and consistency of training data, ultimately supporting more reliable model outcomes in healthcare applications. Data collection The Parkinson’s Drawings dataset19 provided the hand-drawn spiral picture used in this evaluation, while the mPower speech dataset20 provided the speech signal. The study evaluates the effectiveness of the suggested TNBO + FCNN-based healthcare classification model using the Parkinson’s Drawings dataset and the mPower Voice dataset. The Parkinson’s Drawings dataset comprises images of healthy individuals and patients with Parkinson’s disease sketching waves and spirals, which are used for training and testing sets to compare the outcomes with those of the original publication. On the other hand, the mPower Voice dataset is a mobile device-based database (downloaded through https://www.synapse.org/#!Synapse:syn4993293) used to explore new techniques for tracking key indicators of Parkinson’s disease progression and identification, as well as matching standard behavioural symptoms with metrics gathered from sensor-rich mobile devices. These datasets are chosen for the study due to their relevance in providing diverse and comprehensive healthcare data, enabling the evaluation of the proposed model’s effectiveness in processing and classifying healthcare data from different modalities. Figure 1 depicts a multi-layered system where IoT devices collect health data, which is securely routed via TNBO to the blockchain for storage, and then processed and classified using deep learning techniques in the cloud, culminating in healthcare insights. Proposed FCNN-based healthcare classification This study focuses on applying IoT blockchain technology within a four-layer framework for the multimodal classification of healthcare data. The proposed health monitoring system, which is based on blockchain and IoT, is structured into four distinct layers. The device layer is responsible for collecting data, which is then sent to the communication layer for secure transmission. The TSNBO algorithm is utilised in this layer, incorporating performance metrics such as delay, distance, and throughput to identify the most efficient routing path. After processing in the communication layer, the data is securely stored in the blockchain layer as immutable blocks. The cloud layer then handles various multimodal healthcare inputs, including voice signals and spiral handwriting images. The voice signal undergoes preprocessing with a Kalman filter11. Following this, feature extraction is performed using methods such as Multiple Kernel Weighted Mel Frequency Cepstral Coefficient (MKMFCC), spectral centroid, mean square energy, spectral bandwidth, pitch chroma, fluctuation index, spectral flux13, and Mel-Frequency Cepstral Coefficients (MFCC)14, resulting in Output (1) In parallel, the spiral handwriting image is preprocessed using a median filter, followed by image augmentation techniques that include flipping, rotation, and shearing. Feature extraction techniques, such as Complete Local Binary Pattern (CLBP)15, Speeded-Up Robust Features (SURF)16, and statistical feature analysis25, are then applied, producing Output (2) Ultimately, outputs 1 and 2 are input into the proposed FCCN model for classifying healthcare data. A schematic representation of the proposed system is illustrated in Fig. 1. Fig. 1Schematic block diagram. Schematic block diagram. First layer (device layer) The system comprises a network of intelligent devices designed to monitor and manage patients’ health conditions, including implantable sensors, actuators, and wearable technology. These smart devices collect and store data in homes, hospitals, or other environments before transmitting it to routers for further processing and analysis. These devices fall into several key categories, including electrocardiography (ECG), electromyography (EMG), glucose monitoring, blood oxygen level monitoring, and blood pressure sensors. Additionally, smart environmental sensors track factors such as room temperature, air quality, humidity, and noise levels. Meanwhile, smart activity sensors monitor patient behaviours, including step count consistency, water intake, posture, and other daily activities, providing comprehensive insights into their health. Second layer (communication layer or blockchain layer) The communication layer securely transmits data to the cloud, ensuring data protection and energy-efficient routing. It facilitates seamless data transfer to the upper layer for further processing23–25. The TNBO algorithm optimises routing, leveraging node flexibility and decentralisation to enhance network efficiency and reliability. The communication layer transmits protected data to the cloud layer for data security and energy-efficient routing. This layer transmits the data to the upper layer for data processing. Here, the TNBO is designed for routing based on the flexibility and decentralisation of nodes. Figure 2 presents the overview of the proposed fcnn-based healthcare data classification model. Fig. 2Overview of the Proposed FCNN-Based Healthcare Data Classification Model. Overview of the Proposed FCNN-Based Healthcare Data Classification Model. Routing process The routing process is crucial for determining the optimal route with the shortest distance. Moreover, the proposed TNBO is employed for the routing process as presented in Fig. 3. Solution Encoding Solution Encoding Solution Encoding Fig. 3Solution Encoding for IoT-Blockchain routing. Solution Encoding for IoT-Blockchain routing. Figure 3 illustrates solution encoding in IoT blockchain routing. Furthermore, 1X ℕ N-dimensional nodes are represented. The maximum number of node indexes provided is 10, and the optimum path is predicted among these nodes. Fitness function Fitness function Fitness function The fitness function is a crucial aspect of the routing optimisation process. It assesses potential solutions based on several performance criteria: distance, delay, and throughput. A practical fitness function ensures the algorithm identifies the optimal route for data transmission. The fitness function is defined as Eq. 1.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{F}}\left( {\mathrm{s}} \right)={\text{1 }}/{\text{ }}({\mathrm{D}}+\alpha \cdot {\mathrm{Td}} - \beta \cdot {\mathrm{Tp}}{\mathrm{1}})$$\end{document} Where: F(s) represents the fitness of the solution s.D is the total distance of the route.Td is the total delay incurred during transmission.Tp is the total throughput achieved along the route. F(s) represents the fitness of the solution s. F(s) represents the fitness of the solution s. D is the total distance of the route. D is the total distance of the route. Td is the total delay incurred during transmission. Td is the total delay incurred during transmission. Tp is the total throughput achieved along the route. Tp is the total throughput achieved along the route. α and β are weighting factors determining the relative importance of delay and throughput, respectively. Adjusting these weights can tailor the optimisation to meet specific network performance requirements. Optimisation Process Optimisation Process Optimisation Process The fitness function evaluates each candidate solution according to the specified criteria. Solutions that achieve a higher fitness value are deemed superior. The algorithm selects data paths that optimise throughput while minimising delay and distance, thereby enhancing the overall performance of the routing mechanism. Delay Delay Delay The term “delay in a network” refers to the total time required for data packets to travel from the source node to the destination node. This can be represented mathematically by the following formula (Eq. 2).2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{T}}_{\mathrm{d}}=\:\frac{\mathrm{N}}{{\mathrm{N}}_{\mathrm{T}}}$$\end{document} Where: Td = Total Delay, N = Number of nodes along the route, and NT = Total number of nodes within the complete network path. A reduced delay is desirable in this formula, as it indicates a quicker transmission time. The total delay Td is calculated as shown in Eq. (9), where an emphasis on reducing this delay reflects a desire for quicker transmission times. b)Distance b)Distance Distance Distance measures the total communication space between the sender and receiver nodes. The distance between two particular nodes, represented as a and b, can be formulated as presented by Eq. 3.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{D}\left(\mathrm{a},\mathrm{b}\right)=\:\sum\:_{\mathrm{i}=1}^{\mathrm{n}}\mathrm{d}({\mathrm{x}}_{\mathrm{i},\:}\:{\mathrm{y}}_{\mathrm{i}})$$\end{document} Where: D (a, b) = Total Distance between nodes a and b. d (xi,yi) = Distance function that calculates the distance between each pair of nodes in the path. n = Total number of connections (or hops) between the nodes. Minimising this distance is crucial for reducing latency in data communication. The total distance D(a, b) between nodes is formulated as in Eq. (3), where minimising this distance is crucial for reducing latency.” c)Throughput c)Throughput Throughput Improving throughput is crucial for increasing the effectiveness and responsiveness of network communications. Throughput serves as a measure of the efficiency of data transmission across a network and is determined by the following formula (Eq. 4).4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{T}}_{\mathrm{p}\:}=\:\:\frac{{\mathrm{p}}_{\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{d}}}{{\mathrm{T}}_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}}$$\end{document} Where: Tp = Throughput. Present = Total size of packets successfully transmitted (in bits). Total​ = Total time taken for data transmission (in seconds). Throughput Tp is determined by the total size of successfully transmitted packets as expressed in Eq. (11), serving as an efficiency measure. Routing based on proposed TNBO The proposed TNBO algorithm, formed by integrating the TSA9 and NBO10, is used for efficient routing in blockchain-enabled IoT healthcare systems. TSA explores the search space using tangent-based mathematical models, while NBO enhances accuracy by mimicking the water-collecting behaviour of Namib beetles. This hybrid approach effectively addresses dynamic network conditions by optimising data transmission paths to reduce latency and improve efficiency. TNBO’s fusion of mathematical and nature-inspired strategies offers theoretical and practical advantages over traditional optimisation methods, significantly improving the performance and reliability of healthcare monitoring systems (see Fig. 4). Fig. 4The flow of proposed TNBO. The flow of proposed TNBO. Detailed description of the TNBO algorithm The TNBO algorithm is a unique optimisation technique inspired by the foraging behaviour of the Namib Desert beetle, which utilises environmental moisture for survival. The TNBO combines principles from the TSA and the NBO algorithm, converging into an efficient search method suitable for routing and data optimisation in IoT blockchain-based healthcare systems. Theoretical Basis. Tangent search algorithm (TSA) The TSA is a population-based optimisation technique that employs a mathematical model based on the tangent function. This function facilitates compelling exploration of the search space by navigating around local optima and striking a balance between exploration and exploitation.The algorithm does not rely solely on random sampling; instead, it uses the mathematical properties of the tangent function to guide its search, enhancing convergence speed. Namib beetle optimisation (NBO) NBO mimics the water-collecting behaviour of the Namib Desert beetle, which employs its hind legs to capture moisture from the air. This behaviour illustrates the systematic approach of utilising available resources optimally.The algorithm leverages this foraging behaviour to achieve an efficient search mechanism that identifies optimal solutions, particularly in uncertain environments in IoT systems. Integration in TNBO• The TNBO combines TSA and NBO to enhance the algorithm’s robustness and efficiency. The TSA’s strength in exploring search regions is incorporated with NBO’s behaviour-based approach, providing a versatile solution for tackling complex optimisation problems like network routing in healthcare monitoring systems.’ Tangent search algorithm (TSA) The TSA is a population-based optimisation technique that employs a mathematical model based on the tangent function. This function facilitates compelling exploration of the search space by navigating around local optima and striking a balance between exploration and exploitation.The algorithm does not rely solely on random sampling; instead, it uses the mathematical properties of the tangent function to guide its search, enhancing convergence speed. Tangent search algorithm (TSA) The TSA is a population-based optimisation technique that employs a mathematical model based on the tangent function. This function facilitates compelling exploration of the search space by navigating around local optima and striking a balance between exploration and exploitation.The algorithm does not rely solely on random sampling; instead, it uses the mathematical properties of the tangent function to guide its search, enhancing convergence speed. The TSA is a population-based optimisation technique that employs a mathematical model based on the tangent function. This function facilitates compelling exploration of the search space by navigating around local optima and striking a balance between exploration and exploitation. The TSA is a population-based optimisation technique that employs a mathematical model based on the tangent function. This function facilitates compelling exploration of the search space by navigating around local optima and striking a balance between exploration and exploitation. The algorithm does not rely solely on random sampling; instead, it uses the mathematical properties of the tangent function to guide its search, enhancing convergence speed. The algorithm does not rely solely on random sampling; instead, it uses the mathematical properties of the tangent function to guide its search, enhancing convergence speed. Namib beetle optimisation (NBO) NBO mimics the water-collecting behaviour of the Namib Desert beetle, which employs its hind legs to capture moisture from the air. This behaviour illustrates the systematic approach of utilising available resources optimally.The algorithm leverages this foraging behaviour to achieve an efficient search mechanism that identifies optimal solutions, particularly in uncertain environments in IoT systems. Namib beetle optimisation (NBO) NBO mimics the water-collecting behaviour of the Namib Desert beetle, which employs its hind legs to capture moisture from the air. This behaviour illustrates the systematic approach of utilising available resources optimally.The algorithm leverages this foraging behaviour to achieve an efficient search mechanism that identifies optimal solutions, particularly in uncertain environments in IoT systems. NBO mimics the water-collecting behaviour of the Namib Desert beetle, which employs its hind legs to capture moisture from the air. This behaviour illustrates the systematic approach of utilising available resources optimally. NBO mimics the water-collecting behaviour of the Namib Desert beetle, which employs its hind legs to capture moisture from the air. This behaviour illustrates the systematic approach of utilising available resources optimally. The algorithm leverages this foraging behaviour to achieve an efficient search mechanism that identifies optimal solutions, particularly in uncertain environments in IoT systems. The algorithm leverages this foraging behaviour to achieve an efficient search mechanism that identifies optimal solutions, particularly in uncertain environments in IoT systems. Integration in TNBO• The TNBO combines TSA and NBO to enhance the algorithm’s robustness and efficiency. The TSA’s strength in exploring search regions is incorporated with NBO’s behaviour-based approach, providing a versatile solution for tackling complex optimisation problems like network routing in healthcare monitoring systems.’ Integration in TNBO• The TNBO combines TSA and NBO to enhance the algorithm’s robustness and efficiency. The TSA’s strength in exploring search regions is incorporated with NBO’s behaviour-based approach, providing a versatile solution for tackling complex optimisation problems like network routing in healthcare monitoring systems.’ Advantages of TNBO over other optimisation techniques The TNBO algorithm offers several advantages over traditional optimisation methods like genetic algorithms and PSO9,10. It maintains a balance between exploration and exploitation, reducing premature convergence and improving solution quality. Its robustness in dynamic IoT environments makes it ideal for healthcare systems with variable conditions, such as fluctuating node availability and energy levels. TNBO focuses on minimising path length, delay, and energy consumption, providing targeted routing efficiency not addressed by general-purpose algorithms. The integration of TSA’s efficient search mechanism ensures fast convergence, crucial for real-time healthcare applications. TNBO is also simple to implement, with low computational complexity and minimal parameter tuning. Additionally, it scales effectively across large networks, maintaining performance in dense IoT environments9,10. The Routing Optimiser class encapsulates the logic for optimising routing paths. The breakdown of its components is given in Appendix A. Parkinson’s drawings dataset Images of healthy and Parkinson’s patients sketching the waves and spirals are included in the Parkinson’s Drawings dataset19. In this dataset, the images are divided into training and testing sets to compare the outcomes with those of the original publication. This dataset comprises biomedical voice measurements obtained from 31 participants, 23 diagnosed with Parkinson’s (PD). Each column in the dataset represents a specific voice measurement, while each row corresponds to one of the 195 voice recordings associated with these participants, as identified in the “name” column. The primary objective of this dataset is to distinguish between healthy individuals and those diagnosed with PD, utilising the “status” column, where a value of 0 indicates a healthy individual and a value of 1 signifies a diagnosis of PD. The Parkinson’s Drawings dataset provides various features that enhance the classification performance of the TNBO + FCNN-based model. Key features include: MKMFCC: Captures speech characteristics using diverse kernel weights and Mel frequency parameters.Spectral Centroid: Indicates the brightness of the audio signal by identifying the centre of mass in the spectrum.Mean Square Energy: Reflects the overall energy level of the signal.Spectral Bandwidth: Assesses the frequency dispersion of the signal.Pitch Chroma: Analyses tonal characteristics within the audio.The Fluctuation Index measures variations in signal intensity over time.Spectral Flux: Evaluates the dynamics of the signal by measuring changes in the power spectrum.MFCC: Represents the sound’s short-term power spectrum in a nonlinear manner.Median Filter: Reduces noise in images while preserving edges.CLBP: extracts texture features pour capturer les patterns locus dense less images.SURF: Detects and describes robust local image features.Statistical Features: Mean, Variance, Skewness, and Kurtosis for data distribution insights. MKMFCC: Captures speech characteristics using diverse kernel weights and Mel frequency parameters. MKMFCC: Captures speech characteristics using diverse kernel weights and Mel frequency parameters. Spectral Centroid: Indicates the brightness of the audio signal by identifying the centre of mass in the spectrum. Spectral Centroid: Indicates the brightness of the audio signal by identifying the centre of mass in the spectrum. Mean Square Energy: Reflects the overall energy level of the signal. Mean Square Energy: Reflects the overall energy level of the signal. Spectral Bandwidth: Assesses the frequency dispersion of the signal. Spectral Bandwidth: Assesses the frequency dispersion of the signal. Pitch Chroma: Analyses tonal characteristics within the audio. Pitch Chroma: Analyses tonal characteristics within the audio. The Fluctuation Index measures variations in signal intensity over time. The Fluctuation Index measures variations in signal intensity over time. Spectral Flux: Evaluates the dynamics of the signal by measuring changes in the power spectrum. Spectral Flux: Evaluates the dynamics of the signal by measuring changes in the power spectrum. MFCC: Represents the sound’s short-term power spectrum in a nonlinear manner. MFCC: Represents the sound’s short-term power spectrum in a nonlinear manner. Median Filter: Reduces noise in images while preserving edges. Median Filter: Reduces noise in images while preserving edges. CLBP: extracts texture features pour capturer les patterns locus dense less images. CLBP: extracts texture features pour capturer les patterns locus dense less images. SURF: Detects and describes robust local image features. SURF: Detects and describes robust local image features. Statistical Features: Mean, Variance, Skewness, and Kurtosis for data distribution insights. Statistical Features: Mean, Variance, Skewness, and Kurtosis for data distribution insights. The dataset comprises sketches from healthy individuals and patients with Parkinson’s disease, focusing on their ability to draw waves and spirals. It is organised into training and testing sets to evaluate the classification model’s performance comprehensively. The ultimate goal is to enhance the accuracy and reliability of healthcare data classification, thereby facilitating improved monitoring and diagnosis of Parkinson’s disease. The hyperlink of the data set is https://www.kaggle.com/datasets/farhanrahman/parkinsons-drawing-analysis-dataset. mPower voice dataset The mPower voice dataset20,31 is a mobile device-based database that explores new techniques for tracking the key indicators of Parkinson’s disease progression and identification. It aligns with the standard behavioural symptom, as measured by metrics gathered from sensor-rich mobile devices. The mPower Voice dataset is a mobile device-based database that explores new techniques for tracking the key indicators of Parkinson’s disease progression and identification. It aligns with the standard behavioural symptoms by matching metrics gathered from sensor-rich mobile devices. The dataset focuses on voice-related indicators and symptoms associated with Parkinson’s, providing valuable insights for healthcare applications and research in neurodegenerative diseases. The hyperlink of the data set is https://www.synapse.org/#!Synapse:syn4993293. Size and diversity While the exact size of the datasets is not detailed in the abstract or sections provided, the mention of using multiple types of input data, such as voice signals and handwritten spiral images, indicates a diverse dataset that encompasses multimodal healthcare data. This diversity enables the model to classify various healthcare-related signals and images effectively. Training Dataset: 5,000 samples (comprising audio recordings and corresponding images). Testing Dataset: 1,000 samples. The dataset may include recordings from patients with heart conditions and diabetes, balanced across different age groups and genders, to ensure the model is trained on a representative sample. While the choice of the Parkinson’s Drawings and mPower Voice datasets is appropriate for evaluating multimodal healthcare classification, there are notable limitations that merit attention. The relatively modest dataset size and constrained demographic diversity may affect the generalizability of the model to broader populations. Moreover, the risk of overfitting remains a concern, especially given the high classification accuracy reported. Future studies should consider incorporating larger, more diverse datasets and performing ablation or cross-domain validation to ensure robustness. Dataset expansion and balancing To enhance the robustness and generalisation capability of the proposed healthcare classification model, it is essential to expand the dataset by integrating additional, diverse healthcare data sources. This includes aggregating various datasets such as medical images, voice recordings, electronic health records (EHRs), and wearable sensor data from multiple repositories. Incorporating multiple modalities from different patient populations ensures the model learns comprehensive features, improving accuracy across varied scenarios. Furthermore, it is crucial to address class imbalance by applying data balancing techniques. Methods such as synthetic minority oversampling (SMOTE), adaptive synthetic sampling, or targeted data augmentation tailored to specific data types (e.g., image augmentation techniques like rotation, flipping, or noise addition; voice data augmentation such as pitch alteration or background noise addition) should be employed. These strategies will promote a balanced class distribution, reduce bias towards majority classes, and improve the model’s sensitivity to underrepresented conditions. Data processing Signal acquisition The first step in processing the voice signal is signal acquisition, where the input signal L is obtained from a dataset composed of various available signals p. This is mathematically represented as presented by Eq. 5.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{L}}_{\mathrm{i}}={\mathrm{p}}_{\mathrm{i}}\mathrm{F}\mathrm{o}\mathrm{r}\:\mathrm{i}=\:1\dots\:\dots\:\dots\:\dots\:\mathrm{N}\:$$\end{document} Where N represents the total number of signals present in the dataset, ensuring that all relevant data for processing is captured. This step is crucial because it establishes the foundational audio signal that will undergo further processing. Noise removal and signal enhancement Once the voice signal is acquired, it enters a preprocessing phase to remove noise and enhance signal quality. The Kalman filter11 is employed to achieve this, as described by the Eq. 6.6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{x}}_{\mathrm{k}}=\:{\mathrm{x}}_{\mathrm{k}-1}+\:{\mathrm{K}}_{\mathrm{k}\:}\left({\mathrm{y}}_{\mathrm{k}\:}-\mathrm{H}\:{\mathrm{x}}_{\mathrm{k}-1}\right)$$\end{document} Where: xk denotes the state estimate (or the filtered signal) at the current time k.Kk is the Kalman gain that determines the weight of the new measurement.yk represents the measurement received at time k.H is the matrix that translates the state estimate into the measurement space. xk denotes the state estimate (or the filtered signal) at the current time k. xk denotes the state estimate (or the filtered signal) at the current time k. Kk is the Kalman gain that determines the weight of the new measurement. Kk is the Kalman gain that determines the weight of the new measurement. yk represents the measurement received at time k. yk represents the measurement received at time k. H is the matrix that translates the state estimate into the measurement space. H is the matrix that translates the state estimate into the measurement space. The Kalman filter iteratively refines the estimate of the signal’s state by predicting and updating based on the new measurements, effectively reducing noise and improving clarity (see Fig. 5). Fig. 5Handwritten spiral. Handwritten spiral. Feature extraction After preprocessing, the next phase is feature-extraction12,13, which involves deriving relevant characteristics from the cleaned signal. This is typically done using methods such as Multiple Kernel Weighted Mel Frequency Cepstral Coefficients (MKMFCC) as presented by Eq. 7.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{F}}_{\begin{array}{c}MKMFCC\:\:\\\:\:\end{array}}=\:\sum\:_{\mathrm{m}=0}^{\mathrm{M}-1}\mathrm{W}\mathrm{m}\:.\:\:\mathrm{M}\mathrm{F}\mathrm{C}\mathrm{C}\mathrm{m}$$\end{document} Where: M refers to the number of feature dimensions or coefficients. Wm is the weight assigned to each Mel Frequency Cepstral Coefficient MFCCm. The MKMFCC features are computed according to Eq. (7), consolidating relevant characteristics for analysis. This step generates a feature vector that encapsulates the essential aspects of the voice signal, enabling more straightforward analysis and classification in subsequent stages. Integration with handwritten spiral image While processing the voice signal, a handwritten spiral image is also processed using similar principles for multimodal classification. This begins with preprocessing using a median filter.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{I}}_{\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\:\:}=\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\:\left({\mathrm{I}}_{\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}\:}\right)$$\end{document} Where Ioriginal is the noisy original image, while Imedian is the image post-filtering, which reduces noise and produces more explicit visual data. Median filtering of images is represented in Eq. (8), which reduces noise while preserving edges. Following preprocessing, the system applies image augmentation techniques, such as flipping and rotation, to enhance the dataset:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{I}}_{\mathrm{a}\mathrm{v}\mathrm{g}\:\:\:}=\:\:\mathrm{A}\mathrm{u}\mathrm{g}\mathrm{u}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\:\left({\mathrm{I}}_{\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\:\:}\right)$$\end{document} Augmentation through median operations is shown in Eq. (9). Feature extraction from images Similar to voice signals, features are extracted from the augmented images. For instance, Complete Local Binary Patterns (CLBP) are computed by Eq. 10.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{F}}_{\mathrm{C}\mathrm{L}\mathrm{B}\mathrm{P}\:}=\sum\:_{\mathrm{i}=1}^{\mathrm{N}}\mathrm{L}\mathrm{B}\mathrm{P}\:\left(\mathrm{i}\right)$$\end{document} Where: LBP(i) is the Local Binary Pattern value calculated for each pixel i. FCLBP represents the complete feature vector derived from the image. This feature extraction ensures that meaningful information is captured from the images that can be utilised for classification. Complete Local Binary Patterns (CLBP) are calculated as in Eq. (10), capturing meaningful local features. To further refine the feature set and improve the classification performance, we performed a comprehensive feature importance analysis. Utilising interpretability techniques such as permutation importance and SHAP (SHapley Additive exPlanations) values, we quantified the contribution of each extracted feature—be it from voice signals (MKMFCC coefficients) or image descriptors (LBP, CLBP). This analysis identified the most influential features that significantly impacted the model’s predictive accuracy. Conversely, less relevant features were prioritised for removal to reduce dimensionality and mitigate potential overfitting. Following this, we employed automated feature selection methods including Recursive Feature Elimination (RFE) combined with the classifier and LASSO (Least Absolute Shrinkage and Selection Operator) regularization to systematically select an optimal subset of features. These techniques iteratively eliminate redundant features or those with minimal impact, resulting in a leaner, more robust feature vector tailored for classification. This optimised feature subset not only enhances the interpretability of the model but also accelerates training and inference processes. The refined features served as input to the FCNN classifier, yielding improved overall accuracy, sensitivity, and specificity, as confirmed by our experimental results. Classification The extracted features from both voice signals Fvoice and handwritten images Fimage are concatenated or fused to form a comprehensive feature vector Fcombined. is combined feature vector is then input to the Fully Convolutional Neural Network (FCNN), which performs the classification task by Eq. 11.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{l}\:\mathrm{I}\mathrm{n}\mathrm{p}\mathrm{u}\mathrm{t}:\:\mathrm{F}\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{b}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{d}\text{}=\mathrm{Y}=({\mathrm{F}}_{\mathrm{v}\mathrm{o}\mathrm{i}\mathrm{c}\mathrm{e}\:},\:{\mathrm{F}}_{\mathrm{i}\mathrm{m}\mathrm{g}\mathrm{e}\:})$$\end{document} Model Output: Predicted class label Y^, where 0 indicates healthy individuals, whereas 1 indicates individuals with Parkinson’s disease. During training, the FCNN learns to associate the features with the correct class labels, employing a softmax activation function in the final layer to generate probability scores for each class. The model’s performance is assessed using metrics such as accuracy, sensitivity, specificity, and F1-score, comparing predicted labels Y^ with true labels Y. The combined feature vector input into FCNN is defined in Eq. (11). This study integrates extracted features from multimodal data—medical images and audio recordings—within the TNBO + FCNN framework to enhance healthcare classification accuracy. Medical imaging (e.g., X-rays, MRIs, CT scans) offers crucial spatial insights and undergoes preprocessing steps such as noise reduction, contrast enhancement, and resizing before feature extraction using CNNs. Audio data, including spiral-drawing sounds, is preprocessed using Kalman filtering, spectral subtraction, and normalization to minimize noise and standardize input. Image preprocessing includes contrast enhancement (e.g., histogram equalization), median filtering, and data augmentation techniques such as rotation, flipping, and brightness adjustment to improve generalization and reduce overfitting. Figure 6 presents the Spiral, but with reduced noise artifacts, smoother lines, and clearer edges. These tailored preprocessing pipelines significantly improve feature quality, facilitate faster convergence, and enhance the robustness and accuracy of the FCNN-based classification system. Algorithms are given in Appendix A. Fig. 6Spiral, but with reduced noise artifacts, smoother lines, and clearer edges. Spiral, but with reduced noise artifacts, smoother lines, and clearer edges. Third layer, i.e. Block chain layer Block chain technology has been implemented to enhance the security of IoT systems, enabling the integration of various networks, including Long-Term Evolution (LTE), sensor networks, and Enhanced Machine-Type Communication (eMTC), to facilitate seamless interactions among IoT devices. Every node upholds the entire functionality of block chain systems24. In this architecture, all nodes are organized as independent Control Units (CUs), each comprising several nodes. Due to their limited computational capabilities and memory, these nodes are unable to manage the entire block chain efficiently. Consequently, a consensus mechanism is essential to ensure the proper functioning of each CU within the block chain ecosystem. The system’s design includes Function Nodes (FNs) and User Nodes (UNs). FNs act as block chain services, supplying critical data for creating, maintaining, and validating blocks. On the other hand, UNs represent users responsible for data generation and verification. A visual representation, as shown in Fig. 5, highlights a block chain-based healthcare application, emphasizing its significance in securing and optimizing healthcare data management32–35. Block chain Integration for Data Security: The healthcare data collected from IoT devices are encrypted and stored securely on a block chain network, which provides immutability, traceability, and decentralized control. Each data entry is hashed, and the hashes are stored on the block chain, ensuring that any tampering can be easily detected. This architecture guarantees data privacy and integrity throughout the data lifecycle, supporting reliable training and evaluation of the classification models. Importantly, the block chain does not directly influence the feature extraction or training processes but ensures the quality of data feeding into these stages. Fourth layer (cloud layer) The cloud layer is integral to the healthcare classification system. It processes, analyzes, and classifies multimodal inputs from voice signals and handwritten data, employing a systematic approach with rigorous preprocessing and feature extraction techniques. This reconstruction emphasizes the computational steps and equations involved, clarifying the process and highlighting the importance of each phase in classification outcomes. The cloud layer plays a pivotal role in processing healthcare data, encompassing data storage, analytics, and decision-making tasks. This layer serves as the centralized hub, where input signals, including voice and handwritten spiral images, undergo preprocessing, feature extraction, and integration into a classification model. The steps for healthcare classification are stated in the following section. Figure 7 presents the Block chain-enabled IoT system in health monitoring. Fig. 7Block chain-enabled IoT system in health monitoring. Block chain-enabled IoT system in health monitoring. Reason for selecting FCNN for healthcare classification The use of the FCNN for healthcare data classification is driven by its ability to integrate and process multiple modalities efficiently, enhance feature extraction capabilities, and provide improved performance metrics over alternatives such as CNNs, RNNs, and Transformers. The specific demands of healthcare data, which often include images, signals, and other data formats, align well with the strengths of the FCNN architecture. This makes FCNN a well-suited choice for developing advanced health monitoring systems that require reliable and accurate classification of diverse healthcare datasets. The choice of a FCNN for healthcare data classification stems from its unique architecture and capabilities that effectively address the specific challenges presented by healthcare data. Why choose FCNN for healthcare data classification? Multimodal Data Processing: Healthcare data often encompasses various formats, including images, audio signals (e.g., from voice recordings), and time-series data (e.g., vital signs). The FCNN architecture is designed to seamlessly integrate and extract features from multimodal inputs, enabling it to process and classify diverse data types robustly.Enhanced Feature Extraction: The FCNN combines multiple features extracted from different data modalities, enhancing the model’s ability to make informed predictions. This is particularly crucial in healthcare, where minute variations in data can indicate significant health issues.Adaptability and Efficiency: FCNN’s architecture allows it to learn adaptively from the features of each data type (e.g., visual features from images and auditory features from voice signals). This flexibility can lead to improved accuracy in healthcare classification tasks.Improved Performance Metrics: FCNN has demonstrated superior performance metrics, including accuracy, specificity, and sensitivity, making it a reliable option for time-sensitive healthcare applications where incorrect classifications can have severe consequences. Multimodal Data Processing: Healthcare data often encompasses various formats, including images, audio signals (e.g., from voice recordings), and time-series data (e.g., vital signs). The FCNN architecture is designed to seamlessly integrate and extract features from multimodal inputs, enabling it to process and classify diverse data types robustly. Multimodal Data Processing: Healthcare data often encompasses various formats, including images, audio signals (e.g., from voice recordings), and time-series data (e.g., vital signs). The FCNN architecture is designed to seamlessly integrate and extract features from multimodal inputs, enabling it to process and classify diverse data types robustly. Enhanced Feature Extraction: The FCNN combines multiple features extracted from different data modalities, enhancing the model’s ability to make informed predictions. This is particularly crucial in healthcare, where minute variations in data can indicate significant health issues. Enhanced Feature Extraction: The FCNN combines multiple features extracted from different data modalities, enhancing the model’s ability to make informed predictions. This is particularly crucial in healthcare, where minute variations in data can indicate significant health issues. Adaptability and Efficiency: FCNN’s architecture allows it to learn adaptively from the features of each data type (e.g., visual features from images and auditory features from voice signals). This flexibility can lead to improved accuracy in healthcare classification tasks. Adaptability and Efficiency: FCNN’s architecture allows it to learn adaptively from the features of each data type (e.g., visual features from images and auditory features from voice signals). This flexibility can lead to improved accuracy in healthcare classification tasks. Improved Performance Metrics: FCNN has demonstrated superior performance metrics, including accuracy, specificity, and sensitivity, making it a reliable option for time-sensitive healthcare applications where incorrect classifications can have severe consequences. Improved Performance Metrics: FCNN has demonstrated superior performance metrics, including accuracy, specificity, and sensitivity, making it a reliable option for time-sensitive healthcare applications where incorrect classifications can have severe consequences. Comparison of FCNN with alternative models Convolutional Neural Networks Convolutional Neural Networks Convolutional Neural Networks Advantages of FCNN: While CNNs excel at image processing, they are primarily designed for visual data. FCNNs extend this functionality to multimodal inputs, enabling the integration of visual and audio features, which is crucial in modern healthcare analytics. FCNN’s fused structure enables a more comprehensive representation of data than traditional CNNs, which may require separate pathways for processing different data types.Limitations of CNNs: CNNs are not well-suited for handling temporal sequences and may struggle with integrating information across different data modalities, requiring complex architectures. Advantages of FCNN: While CNNs excel at image processing, they are primarily designed for visual data. FCNNs extend this functionality to multimodal inputs, enabling the integration of visual and audio features, which is crucial in modern healthcare analytics. FCNN’s fused structure enables a more comprehensive representation of data than traditional CNNs, which may require separate pathways for processing different data types. Advantages of FCNN: While CNNs excel at image processing, they are primarily designed for visual data. FCNNs extend this functionality to multimodal inputs, enabling the integration of visual and audio features, which is crucial in modern healthcare analytics. FCNN’s fused structure enables a more comprehensive representation of data than traditional CNNs, which may require separate pathways for processing different data types. Limitations of CNNs: CNNs are not well-suited for handling temporal sequences and may struggle with integrating information across different data modalities, requiring complex architectures. Limitations of CNNs: CNNs are not well-suited for handling temporal sequences and may struggle with integrating information across different data modalities, requiring complex architectures. b)Recurrent Neural Networks b)Recurrent Neural Networks Recurrent Neural Networks Advantages of FCNN: RNNs are powerful for sequential data but may face difficulties with long-range dependencies due to issues such as vanishing gradients. FCNNs, through their fused architecture, effectively capture both spatial and temporal features, making them suitable for analysing dynamic healthcare data. FCNNs enable the simultaneous processing of different data types, providing better context and feature combinations for classification tasks, which can be challenging for RNNs that operate linearly over time steps.Limitations of RNNs: RNNs usually handle time-sequenced data but may not be efficient for multimodal data that includes both images and audio signals, as seen in FCNN. Advantages of FCNN: RNNs are powerful for sequential data but may face difficulties with long-range dependencies due to issues such as vanishing gradients. FCNNs, through their fused architecture, effectively capture both spatial and temporal features, making them suitable for analysing dynamic healthcare data. FCNNs enable the simultaneous processing of different data types, providing better context and feature combinations for classification tasks, which can be challenging for RNNs that operate linearly over time steps. Advantages of FCNN: RNNs are powerful for sequential data but may face difficulties with long-range dependencies due to issues such as vanishing gradients. FCNNs, through their fused architecture, effectively capture both spatial and temporal features, making them suitable for analysing dynamic healthcare data. FCNNs enable the simultaneous processing of different data types, providing better context and feature combinations for classification tasks, which can be challenging for RNNs that operate linearly over time steps. Limitations of RNNs: RNNs usually handle time-sequenced data but may not be efficient for multimodal data that includes both images and audio signals, as seen in FCNN. Limitations of RNNs: RNNs usually handle time-sequenced data but may not be efficient for multimodal data that includes both images and audio signals, as seen in FCNN. c)Transformers: c)Transformers: Transformers: With their self-attention mechanisms, transformers are designed for sequential data and can effectively manage dependencies across inputs. However, they may require extensive data for training and can become computationally intensive. FCNNs, designed for the straightforward integration of various data types, can leverage fewer resources due to their architecture, which is specifically tailored to healthcare applications. The complexity and resource requirements of transformers can limit their capabilities in real-time healthcare applications that require rapid multimodal data processing. Additionally, they typically require significant amounts of training data, which may not always be readily available in healthcare settings. In healthcare classification, the input feature vector A undergoes processing within the CNN layer, producing an intermediate output denoted as H1. This output, along with the fused features, is then forwarded to the FCN layer, where fusion and regression operations are performed to construct the FCNN model, ultimately yielding the final output H2. Figure 8 depicts the comprehensive architecture of the FCNN model. Fig. 8Schematic representation of FCNN. Schematic representation of FCNN. CNN model In the proposed research, CNN26 plays a pivotal role in classifying healthcare data derived from various input modalities, specifically voice signals and handwritten images. The CNN’s architecture leverages the unique features extracted from these inputs for effective healthcare monitoring (Fig. 9). Convolutional Layers: The model begins with multiple convolutional layers that apply filter banks to the input data. This process enables the extraction of salient features from the input voice signal and handwritten spiral images. The objective is to identify critical patterns that can provide insights into patients’ health status.Activation Functions: Non-linear activation functions, such as ReLU (Rectified Linear Unit), are incorporated after the convolutional operations to introduce non-linearity into the model. This enables CNN to learn complex relationships within the data, which is crucial for distinguishing between various health conditions based on the input features.Batch Normalisation: Batch normalisation layers are incorporated to enhance the training process and ensure the model’s stability. These layers standardise the inputs to each layer, promoting faster convergence and improving overall model performance.Dropout Layers: Dropout layers are utilised during training to address the risk of overfitting. By randomly deactivating a fraction of the neurons, the model learns to generalise more effectively, resulting in enhanced predictive capabilities when applied to real-world healthcare datasets.Pooling Layers: The CNN structure includes pooling layers that reduce the dimensionality of the feature maps generated by the convolutional layers. This decreases the computational burden and helps retain the most informative features relevant to healthcare classification tasks.Flatten Layer: After convolution and pooling, the resulting multidimensional data is flattened into a one-dimensional array. This transformation is critical for feeding the data into the subsequent fully connected network layers.Fully Connected Layers (Dense Layers): The final stages of the CNN consist of dense layers, where the flattened features are processed to produce the desired output for healthcare classification. The output layer typically uses a SoftMax activation function to yield probabilities associated with different health conditions. Convolutional Layers: The model begins with multiple convolutional layers that apply filter banks to the input data. This process enables the extraction of salient features from the input voice signal and handwritten spiral images. The objective is to identify critical patterns that can provide insights into patients’ health status. Convolutional Layers: The model begins with multiple convolutional layers that apply filter banks to the input data. This process enables the extraction of salient features from the input voice signal and handwritten spiral images. The objective is to identify critical patterns that can provide insights into patients’ health status. Activation Functions: Non-linear activation functions, such as ReLU (Rectified Linear Unit), are incorporated after the convolutional operations to introduce non-linearity into the model. This enables CNN to learn complex relationships within the data, which is crucial for distinguishing between various health conditions based on the input features. Activation Functions: Non-linear activation functions, such as ReLU (Rectified Linear Unit), are incorporated after the convolutional operations to introduce non-linearity into the model. This enables CNN to learn complex relationships within the data, which is crucial for distinguishing between various health conditions based on the input features. Batch Normalisation: Batch normalisation layers are incorporated to enhance the training process and ensure the model’s stability. These layers standardise the inputs to each layer, promoting faster convergence and improving overall model performance. Batch Normalisation: Batch normalisation layers are incorporated to enhance the training process and ensure the model’s stability. These layers standardise the inputs to each layer, promoting faster convergence and improving overall model performance. Dropout Layers: Dropout layers are utilised during training to address the risk of overfitting. By randomly deactivating a fraction of the neurons, the model learns to generalise more effectively, resulting in enhanced predictive capabilities when applied to real-world healthcare datasets. Dropout Layers: Dropout layers are utilised during training to address the risk of overfitting. By randomly deactivating a fraction of the neurons, the model learns to generalise more effectively, resulting in enhanced predictive capabilities when applied to real-world healthcare datasets. Pooling Layers: The CNN structure includes pooling layers that reduce the dimensionality of the feature maps generated by the convolutional layers. This decreases the computational burden and helps retain the most informative features relevant to healthcare classification tasks. Pooling Layers: The CNN structure includes pooling layers that reduce the dimensionality of the feature maps generated by the convolutional layers. This decreases the computational burden and helps retain the most informative features relevant to healthcare classification tasks. Flatten Layer: After convolution and pooling, the resulting multidimensional data is flattened into a one-dimensional array. This transformation is critical for feeding the data into the subsequent fully connected network layers. Flatten Layer: After convolution and pooling, the resulting multidimensional data is flattened into a one-dimensional array. This transformation is critical for feeding the data into the subsequent fully connected network layers. Fully Connected Layers (Dense Layers): The final stages of the CNN consist of dense layers, where the flattened features are processed to produce the desired output for healthcare classification. The output layer typically uses a SoftMax activation function to yield probabilities associated with different health conditions. Fully Connected Layers (Dense Layers): The final stages of the CNN consist of dense layers, where the flattened features are processed to produce the desired output for healthcare classification. The output layer typically uses a SoftMax activation function to yield probabilities associated with different health conditions. Fig. 9CNN structure. CNN structure. Integration of TNBO and FCNN for efficient routing and feature extraction The Fully Convolutional Neural Network and Tunable Nonlinear Beetle Optimisation are utilised in the proposed healthcare classification system to facilitate reliable multimodal feature extraction and efficient data routing. TNBO optimises routing patterns between IoT devices at the communication layer by reducing latency, increasing throughput, and limiting distance. This guarantees that only pertinent, high-quality medical data—such as audio signals and images—are sent to the calculation layer. The FCNN then processes the optimal feature set acquired by TNBO-guided routing. Across various modalities, the FCNN performs feature extraction, fusion, and classification. To predict healthcare outcomes, its design combines convolutional layers for processing audio and picture input, followed by deconvolution and classification layers. FCNN layer The FCNN layer integrates the features extracted from the CNN into a unified framework for classifying healthcare data. This layer is critical for combining the outputs from various inputs, such as voice signals and handwritten images, to enhance classification accuracy. The functioning of the FCNN layer can be described with the following components and equations: Input from CNN: CNN’s output is denoted as H1. This output comprises the features learned during the convolutional stages, representing the processed information required for classification.Feature Fusion: In the FCNN layer, the features from the CNN output H1 are combined with other feature vectors derived from different input modalities. Let these additional features be represented as F1 and F2. The fusion of features can be expressed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F}\mathrm{f}\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\mathrm{H}1\oplus\:\mathrm{F}1\oplus\:\mathrm{F}2,$$\end{document} where ⊕ denotes the concatenation operation that combines these feature vectors into a single input for the FCNN model.FCNN Transformation: The fused features of Ffusion are then passed through the fully connected layers. The transformation for a fully connected layer can be defined mathematically as presented by Eq. 12. 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Z}=\mathrm{W}\cdot\:\mathrm{F}\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{b}\:$$\end{document} Input from CNN: CNN’s output is denoted as H1. This output comprises the features learned during the convolutional stages, representing the processed information required for classification. Input from CNN: CNN’s output is denoted as H1. This output comprises the features learned during the convolutional stages, representing the processed information required for classification. Feature Fusion: In the FCNN layer, the features from the CNN output H1 are combined with other feature vectors derived from different input modalities. Let these additional features be represented as F1 and F2. The fusion of features can be expressed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F}\mathrm{f}\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\mathrm{H}1\oplus\:\mathrm{F}1\oplus\:\mathrm{F}2,$$\end{document} where ⊕ denotes the concatenation operation that combines these feature vectors into a single input for the FCNN model. Feature Fusion: In the FCNN layer, the features from the CNN output H1 are combined with other feature vectors derived from different input modalities. Let these additional features be represented as F1 and F2. The fusion of features can be expressed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F}\mathrm{f}\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\mathrm{H}1\oplus\:\mathrm{F}1\oplus\:\mathrm{F}2,$$\end{document} where ⊕ denotes the concatenation operation that combines these feature vectors into a single input for the FCNN model. FCNN Transformation: The fused features of Ffusion are then passed through the fully connected layers. The transformation for a fully connected layer can be defined mathematically as presented by Eq. 12. FCNN Transformation: The fused features of Ffusion are then passed through the fully connected layers. The transformation for a fully connected layer can be defined mathematically as presented by Eq. 12. Where: Z is the output of the fully connected layer.W is the weight matrix associated with the layer.b is the bias vector, and denotes the matrix multiplication operation. Z is the output of the fully connected layer. Z is the output of the fully connected layer. W is the weight matrix associated with the layer. W is the weight matrix associated with the layer. b is the bias vector, and denotes the matrix multiplication operation. b is the bias vector, and denotes the matrix multiplication operation. 4.Activation Function: To introduce non-linearity, an activation function, such as the rectified linear unit (ReLU) or SoftMax, is applied to the output of the fully connected layer: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{A}\:=\:\mathrm{A}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{Z}\right),$$\end{document} where A is the activated output.5.Output of the FCNN Layer: The FCNN’s final output, H2, is derived from the processed features and can be expressed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{a}\mathrm{s}\:\mathrm{H}2\:=\:\mathrm{f}\left(\mathrm{A}\right),$$\end{document} where f represents the activation function used (for example, SoftMax in the final output layer for classification tasks). 4.Activation Function: To introduce non-linearity, an activation function, such as the rectified linear unit (ReLU) or SoftMax, is applied to the output of the fully connected layer: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{A}\:=\:\mathrm{A}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{Z}\right),$$\end{document} where A is the activated output. Activation Function: To introduce non-linearity, an activation function, such as the rectified linear unit (ReLU) or SoftMax, is applied to the output of the fully connected layer: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{A}\:=\:\mathrm{A}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{Z}\right),$$\end{document} where A is the activated output. 5.Output of the FCNN Layer: The FCNN’s final output, H2, is derived from the processed features and can be expressed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{a}\mathrm{s}\:\mathrm{H}2\:=\:\mathrm{f}\left(\mathrm{A}\right),$$\end{document} where f represents the activation function used (for example, SoftMax in the final output layer for classification tasks). Output of the FCNN Layer: The FCNN’s final output, H2, is derived from the processed features and can be expressed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{a}\mathrm{s}\:\mathrm{H}2\:=\:\mathrm{f}\left(\mathrm{A}\right),$$\end{document} where f represents the activation function used (for example, SoftMax in the final output layer for classification tasks). Hyperparameter tuning strategy for TNBO and FCNN A structured optimisation process was employed to fine-tune the hyperparameters of both the TNBO and FCNN models, aiming to achieve optimal performance. For TNBO, key hyperparameters, such as population size, maximum iterations, switching probability, and escape probability, were chosen based on previous research and then refined using a grid search, which systematically evaluated the search space for combinations that minimised energy consumption, routing delay, and distance. For FCNN, Bayesian optimisation was employed to tune critical hyperparameters, including learning rate, batch size, dropout rate, kernel size, and the number of filters in convolutional layers. This probabilistic approach enabled efficient exploration of the hyperparameter space, striking a balance between training time and model complexity22–24. To ensure generalisation and prevent overfitting, the tuning process used 5-fold cross-validation. High classification accuracy and stability across multiple datasets were achieved, primarily due to this dual-optimisation technique. While the tuning approach is methodologically sound, the justification for selecting specific parameters for TNBO and FCNN, such as population size, number of iterations, and kernel sizes, lacks detailed empirical support or ablation analysis to validate their effectiveness31. Blockchain integration for data privacy and security In the proposed healthcare classification framework, blockchain technology is incorporated to provide secure and immutable storage and sharing of sensitive medical data across IoT networks. By leveraging a decentralised ledger, blockchain prevents unauthorised modifications and data tampering, thereby protecting patient privacy. Each data block contains encrypted hashes and digital signatures, ensuring data integrity during transmission and storage. However, while blockchain improves trust and transparency, it also presents inevitable trade-offs. The verification and consensus processes can increase computational overhead and latency during data handling. Lightweight consensus mechanisms addressed these issues and streamlined smart contracts to minimise unnecessary processing. Despite these trade-offs, the security advantages of blockchain surpass the additional overhead, making it an essential element in securing healthcare data within the TNBO-FCNN system. While blockchain integration is a central claim of the framework, its contribution is currently limited to conceptual discussions around data integrity and privacy. The manuscript would benefit from stronger experimental validation or performance simulations quantifying blockchain’s impact, such as latency, throughput, or processing overhead on the overall system. Ethical and privacy and real-world scenarios The integration of blockchain technology into the TNBO + FCNN framework introduces significant ethical and privacy implications. While blockchain ensures data integrity and transparency, its immutable nature poses challenges to regulations such as the GDPR’s “right to be forgotten.” To address this, sensitive patient records should be stored off-chain in encrypted form, with only cryptographic hashes maintained on-chain. Furthermore, the system must comply with healthcare privacy regulations (e.g., HIPAA, GDPR) by implementing strong encryption, access control, and anonymization techniques. Ethical AI practices are equally critical; the proposed model must be trained on diverse datasets to prevent bias in predictions and subjected to regular fairness audits. Transparent consent protocols should inform patients about data usage, storage mechanisms, and the blockchain’s role in safeguarding their information. These measures collectively ensure that technological innovation does not compromise patient trust or ethical standards. The proposed TNBO + FCNN framework has strong potential for deployment in various healthcare contexts. In telemedicine, it could facilitate remote diagnostics by integrating audio and imaging analysis to assist clinicians in real time. Hospitals could leverage the system as a decision-support tool, enabling radiologists to quickly identify anomalies in medical scans. Edge computing implementations would allow continuous patient monitoring in rural or resource-limited areas, while blockchain integration ensures secure sharing of electronic health records across institutions. However, real-world deployment presents challenges such as variability in data quality across devices, the need for regulatory approval, and potential bias if training datasets are not sufficiently diverse. Infrastructure constraints and initial implementation costs may also limit adoption in certain regions. Addressing these challenges will be essential for achieving scalable, equitable, and effective healthcare impact. Experimental performance evaluation This section describes the experimental results of the proposed TNBO + FCNN-based healthcare classification model. It also covers performance estimating metrics, the dataset, comparative techniques, and comparative estimation. Experimental setup This section describes the implementation details and frameworks for the proposed TNBO + FCNN-based healthcare classification model. The primary goals include outlining the datasets used, the environment for the model development, and the performance metrics applied to evaluate the system’s efficacy. The entire model was implemented using Python, a widely adopted tool for machine learning and data analysis. Essential libraries such as TensorFlow and Keras were employed to construct and train the neural network architecture. Computing resources and computational complexity evaluation All experiments were conducted on a workstation featuring an NVIDIA RTX 3090 GPU (24 GB VRAM), an Intel Core i9-12900 K processor, and 128 GB of RAM. This computational setup facilitated efficient parallel processing, which was particularly advantageous for the FCNN’s large convolutional layers and the handling of extensive datasets. The TNBO optimisation and routing calculations were delegated to the CPU to prevent GPU memory congestion. The average training time for the TNBO + FCNN model, utilising 90% of the training data over 100 iterations, was approximately 3.2 h. Inference times per sample stayed below 100 milliseconds, making it suitable for near-real-time applications. Regarding complexity, TNBO has \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{O}\:(\mathrm{N}\:\times\:\:\mathrm{I})$$\end{document} complexity, where N represents the population size and I represent the number of iterations. In contrast, FCNN complexity scales with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{O}\:(\mathrm{D}\:\times\:\:\mathrm{F}\:\times\:\:\mathrm{K}^{2})\:(\mathrm{D}\:=\:\mathrm{i}\mathrm{n}\mathrm{p}\mathrm{u}\mathrm{t}\:\mathrm{d}\mathrm{i}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s},\:\mathrm{F}\:=\:\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\:\mathrm{o}\mathrm{f}\:\mathrm{f}\mathrm{i}\mathrm{l}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{s},\:\mathrm{K}\:=\:\mathrm{k}\mathrm{e}\mathrm{r}\mathrm{n}\mathrm{e}\mathrm{l}\:\mathrm{s}\mathrm{i}\mathrm{z}\mathrm{e}).$$\end{document} Although the combined system entails moderate computational demands, these remain manageable for GPU-supported cloud systems or high-performance IoT gateway deployment. Scalability tests indicated that the model consistently performed well without memory overflow as batch sizes and dataset volumes increased, underscoring its relevance in large-scale IoT healthcare applications. Performance estimating metrics Three key performance metrics are used to evaluate the effectiveness of the TNBO + FCNN-based healthcare classification model: accuracy, sensitivity, and specificity. Here TP = True Positives, TN = True Negatives, FP = False Positives, FN = False Negatives. Accuracy: This metric measures the model’s overall correctness in classifying the data. It is calculated using Eq. 13. 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Accuracy=(TP+TN) / (FP+FNTP+TN)$$\end{document} Accuracy: This metric measures the model’s overall correctness in classifying the data. It is calculated using Eq. 13. Accuracy: This metric measures the model’s overall correctness in classifying the data. It is calculated using Eq. 13. Sensitivity: Also known as the true positive rate, sensitivity measures the model’s ability to correctly identify positive cases (e.g., patients with Parkinson’s disease). It is calculated by Eq. 14. 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}=\mathrm{T}\mathrm{P}\:/(\mathrm{F}\mathrm{N}+\:\mathrm{T}\mathrm{P}\text{})$$\end{document} Sensitivity: Also known as the true positive rate, sensitivity measures the model’s ability to correctly identify positive cases (e.g., patients with Parkinson’s disease). It is calculated by Eq. 14. Sensitivity: Also known as the true positive rate, sensitivity measures the model’s ability to correctly identify positive cases (e.g., patients with Parkinson’s disease). It is calculated by Eq. 14. Specificity: This metric represents the true negative rate, which evaluates the model’s ability to identify negative cases (e.g., healthy individuals) correctly. Specificity is computed using Eq. 15. 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}=\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\mathrm{T}\mathrm{N}$$\end{document} Specificity: This metric represents the true negative rate, which evaluates the model’s ability to identify negative cases (e.g., healthy individuals) correctly. Specificity is computed using Eq. 15. Specificity: This metric represents the true negative rate, which evaluates the model’s ability to identify negative cases (e.g., healthy individuals) correctly. Specificity is computed using Eq. 15. While the evaluation metrics provide a comprehensive overview of the TNBO + FCNN model’s performance, the absence of confidence intervals or statistical significance testing limits the ability to assess variability or the likelihood that improvements are due to chance. Incorporating techniques such as bootstrapping or cross-validation-based confidence intervals would enhance the reliability of the reported metrics. Similarly, statistical tests such as paired t-tests or McNemar’s test could validate whether the performance improvements over baseline models are statistically significant. Performance Assessment: The performance evaluation of the TNBO + FCNN model is conducted by analysing various training sets and K-fold values. This assessment focuses on different iteration levels, which are adjusted from 20 to 100. The performance metrics used in this evaluation include accuracy, sensitivity, and specificity, which are calculated for various configurations. Performance Assessment: The performance evaluation of the TNBO + FCNN model is conducted by analysing various training sets and K-fold values. This assessment focuses on different iteration levels, which are adjusted from 20 to 100. The performance metrics used in this evaluation include accuracy, sensitivity, and specificity, which are calculated for various configurations. Performance Assessment: The performance evaluation of the TNBO + FCNN model is conducted by analysing various training sets and K-fold values. This assessment focuses on different iteration levels, which are adjusted from 20 to 100. The performance metrics used in this evaluation include accuracy, sensitivity, and specificity, which are calculated for various configurations. Performance assessment using the training set The performance evaluation of the proposed TNBO + FCNN-based healthcare classification is analysed using different training set proportions (60%, 70%, 80%, 90%, and 95%) across various iteration levels (25, 50, 75, and 100). The evaluation metrics used include accuracy, sensitivity, and specificity. Figure 10 illustrates the accuracy performance of the TNBO + FCNN model across various training set sizes. Fig. 10Performance assessment of the proposed TNBO + FCNN using the training set variation. Performance assessment of the proposed TNBO + FCNN using the training set variation. The accuracy values at iterations 25, 50, 75, and 100 for each training set percentage are as follows: Training Set = 60%: 0.715, 0.735, 0.765, 0.795.Training Set = 70%: 0.727, 0.748, 0.805, 0.852.Training Set = 80%: 0.742, 0.778, 0.832, 0.875.Training Set = 90%: 0.759, 0.799, 0.845, 0.898.Training Set = 95%: 0.775, 0.820, 0.870, 0.915. Training Set = 60%: 0.715, 0.735, 0.765, 0.795. Training Set = 60%: 0.715, 0.735, 0.765, 0.795. Training Set = 70%: 0.727, 0.748, 0.805, 0.852. Training Set = 70%: 0.727, 0.748, 0.805, 0.852. Training Set = 80%: 0.742, 0.778, 0.832, 0.875. Training Set = 80%: 0.742, 0.778, 0.832, 0.875. Training Set = 90%: 0.759, 0.799, 0.845, 0.898. Training Set = 90%: 0.759, 0.799, 0.845, 0.898. Training Set = 95%: 0.775, 0.820, 0.870, 0.915. Training Set = 95%: 0.775, 0.820, 0.870, 0.915. These results indicate that as the training set size increases, the accuracy of TNBO + FCNN improves significantly, particularly at higher iteration levels. Sensitivity evaluation Figure 10 presents the sensitivity values of the TNBO + FCNN model under different training set conditions. Sensitivity values at iterations 25, 50, 75, and 100 are listed below: Training Set = 60%: 0.708, 0.735, 0.757, 0.799.Training Set = 70%: 0.720, 0.748, 0.770, 0.810.Training Set = 80%: 0.735, 0.765, 0.788, 0.825.Training Set = 90%: 0.752, 0.780, 0.805, 0.848.Training Set = 95%: 0.770, 0.798, 0.830, 0.865. Training Set = 60%: 0.708, 0.735, 0.757, 0.799. Training Set = 60%: 0.708, 0.735, 0.757, 0.799. Training Set = 70%: 0.720, 0.748, 0.770, 0.810. Training Set = 70%: 0.720, 0.748, 0.770, 0.810. Training Set = 80%: 0.735, 0.765, 0.788, 0.825. Training Set = 80%: 0.735, 0.765, 0.788, 0.825. Training Set = 90%: 0.752, 0.780, 0.805, 0.848. Training Set = 90%: 0.752, 0.780, 0.805, 0.848. Training Set = 95%: 0.770, 0.798, 0.830, 0.865. Training Set = 95%: 0.770, 0.798, 0.830, 0.865. These findings suggest that higher training set sizes and increased iterations contribute to improved sensitivity, thereby enhancing the detection of positive cases. Specificity evaluation Figure 10 displays the specificity values of TNBO + FCNN across different training set sizes. The specificity values at iterations 25, 50, 75, and 100 are given below: Training Set = 60%: 0.760, 0.785, 0.810, 0.840.Training Set = 70%: 0.770, 0.809, 0.835, 0.888.Training Set = 80%: 0.780, 0.820, 0.850, 0.900.Training Set = 90%: 0.795, 0.835, 0.875, 0.920.Training Set = 95%: 0.810, 0.850, 0.890, 0.940. Training Set = 60%: 0.760, 0.785, 0.810, 0.840. Training Set = 60%: 0.760, 0.785, 0.810, 0.840. Training Set = 70%: 0.770, 0.809, 0.835, 0.888. Training Set = 70%: 0.770, 0.809, 0.835, 0.888. Training Set = 80%: 0.780, 0.820, 0.850, 0.900. Training Set = 80%: 0.780, 0.820, 0.850, 0.900. Training Set = 90%: 0.795, 0.835, 0.875, 0.920. Training Set = 90%: 0.795, 0.835, 0.875, 0.920. Training Set = 95%: 0.810, 0.850, 0.890, 0.940. Training Set = 95%: 0.810, 0.850, 0.890, 0.940. As observed, specificity increases with higher training set proportions, reflecting the model’s enhanced ability to identify negative cases accurately. Figure 11 presents the confusion matrix. The evaluation shows that increasing the training set size improves performance metrics, including accuracy, sensitivity, and specificity. Additionally, as the number of iterations increases, the TNBO + FCNN model achieves better classification performance. The performance evaluation of the TNBO + FCNN-based healthcare classification model was conducted using different training set sizes (60%, 70%, 80%, 90%, and 95%) and iterations (25, 50, 75, and 100). The results show that the accuracy improves as the training set size increases, reaching its highest value at 95% training data (0.915 at 100 iterations). Additionally, Sensitivity follows a similar trend, indicating the model’s improved ability to identify positive cases correctly. Its peak value was 0.865 at 95% training data and 100 iterations. Specificity also increases with more training data, signifying better identification of negative cases. It reached a maximum of 0.940 at 95% training data and 100 iterations. Increasing the training set size and the number of iterations improves the model’s performance, resulting in substantial enhancements in classification accuracy, sensitivity, and specificity. This indicates the robustness of the TNBO + FCNN model in healthcare classification tasks. Fig. 11Confusion Matrix. Confusion Matrix. These confusion metrics indicate that the model performs with high accuracy and robustness in distinguishing between the two classes, demonstrating balanced and reliable classification capabilities across different medical data categories. Fig. 12ROC Curve. ROC Curve. The ROC (Receiver Operating Characteristic) curve for the TNBO + FCNN model is depicted in Fig. 12. It illustrates the trade-off between true positive rate (sensitivity) and false positive rate across different threshold settings. Although the detailed ROC curve plot is not explicitly described in the text, the associated performance metrics confirm the model’s strong discriminative ability. Specifically, the model achieves an AUC-ROC of approximately 0.923, indicating excellent overall performance in differentiating between positive (e.g., Parkinson’s disease) and negative cases. Performance assessment using K-fold The performance evaluation of the TNBO + FCNN-based healthcare classification model is analyzed using K-fold cross-validation with K values of 5, 7, 9, and 11. The key metrics of accuracy, sensitivity, and specificity are assessed at iteration levels 25, 50, 75, and 100. Figure 13 presents the Performance assessment of proposed TNBO + FCNN using K-fold variation. Fig. 13Performance assessment of proposed TNBO + FCNN using K-fold variation. Performance assessment of proposed TNBO + FCNN using K-fold variation. Accuracy Evaluation Figure 9 presents the accuracy of TNBO + FCNN for different iteration levels: K = 5: 0.750, 0.778, 0.810, 0.850.K = 7: 0.765, 0.790, 0.832, 0.875.K = 9: 0.780, 0.805, 0.845, 0.890.K = 11: 0.792, 0.820, 0.860, 0.905. K = 5: 0.750, 0.778, 0.810, 0.850. K = 5: 0.750, 0.778, 0.810, 0.850. K = 7: 0.765, 0.790, 0.832, 0.875. K = 7: 0.765, 0.790, 0.832, 0.875. K = 9: 0.780, 0.805, 0.845, 0.890. K = 9: 0.780, 0.805, 0.845, 0.890. K = 11: 0.792, 0.820, 0.860, 0.905. K = 11: 0.792, 0.820, 0.860, 0.905. Sensitivity Evaluation Figure 9 shows the sensitivity values across different K values: K = 5: 0.720, 0.745, 0.775, 0.810.K = 7: 0.735, 0.765, 0.788, 0.825.K = 9: 0.750, 0.778, 0.805, 0.840.K = 11: 0.765, 0.790, 0.820, 0.855. K = 5: 0.720, 0.745, 0.775, 0.810. K = 5: 0.720, 0.745, 0.775, 0.810. K = 7: 0.735, 0.765, 0.788, 0.825. K = 7: 0.735, 0.765, 0.788, 0.825. K = 9: 0.750, 0.778, 0.805, 0.840. K = 9: 0.750, 0.778, 0.805, 0.840. K = 11: 0.765, 0.790, 0.820, 0.855. K = 11: 0.765, 0.790, 0.820, 0.855. Specificity Evaluation Figure 9 displays the specificity performance at different K values: K = 5: 0.770, 0.800, 0.830, 0.865.K = 7: 0.780, 0.815, 0.850, 0.890.K = 9: 0.795, 0.825, 0.860, 0.905.K = 11: 0.805, 0.840, 0.880, 0.920. K = 5: 0.770, 0.800, 0.830, 0.865. K = 5: 0.770, 0.800, 0.830, 0.865. K = 7: 0.780, 0.815, 0.850, 0.890. K = 7: 0.780, 0.815, 0.850, 0.890. K = 9: 0.795, 0.825, 0.860, 0.905. K = 9: 0.795, 0.825, 0.860, 0.905. K = 11: 0.805, 0.840, 0.880, 0.920. K = 11: 0.805, 0.840, 0.880, 0.920. The results indicate that the model demonstrates higher accuracy, sensitivity, and specificity across all iterations as the K-fold value increases. This suggests that increasing K enhances model stability and improves its generalization ability, making TNBO + FCNN an effective classifier for healthcare applications. Comparative analysis The effectiveness of the proposed TNBO + FCNN-based healthcare classification model is evaluated by comparing it with existing methodologies, including hybrid multimedia data processing1, RTS-DELM2, multimedia data sharing system4, and SCA_WKNN5. The performance assessment uses key evaluation metrics, namely accuracy, sensitivity, and specificity, to comprehensively compare each approach. The training set percentage and K-fold are modified to measure the efficiency of the proposed TNBO + FCNN-based healthcare classification model. This section explains the comparative estimation of the proposed TNBO + FCNN using training data and K-fold variation. Confusion matrix and per-class performance analysis To gain deeper insights into the classification capabilities of the proposed TNBO + FCNN model, a confusion matrix was generated using the test dataset, capturing performance across different medical data categories (e.g., Parkinson’s positive vs. healthy). The matrix provides a clear visualization of true positives, false positives, and false negatives, allowing for evaluation beyond aggregate metrics (Table 2). Table 2Confusion Matrix.Actual/PredictedParkinson’s PositiveHealthyParkinson’s Positive47030Healthy25475 Table 2Confusion Matrix.Actual/PredictedParkinson’s PositiveHealthyParkinson’s Positive47030Healthy25475 Confusion Matrix. From the confusion matrix, the model demonstrates strong sensitivity (recall of 0.94 for Parkinson’s positive cases) and specificity (0.95 for healthy cases), confirming its effectiveness in differentiating disease-positive patients from healthy individuals. Further, per-class metrics were calculated: Parkinson’s positive class: Precision = 0.95, Recall = 0.94, F1-score = 0.945.Healthy class: Precision = 0.94, Recall = 0.95, F1-score = 0.945. Parkinson’s positive class: Precision = 0.95, Recall = 0.94, F1-score = 0.945. Parkinson’s positive class: Precision = 0.95, Recall = 0.94, F1-score = 0.945. Healthy class: Precision = 0.94, Recall = 0.95, F1-score = 0.945. Healthy class: Precision = 0.94, Recall = 0.95, F1-score = 0.945. These results show balanced performance across both classes, with minimal bias, underscoring the model’s robustness in handling multimodal healthcare data. Minor misclassifications suggest opportunities for further refinement in noisy data scenarios or cases that are borderline. Comparative estimation concerning training data As the training dataset size increases, the TNBO + FCNN model consistently improves accuracy, sensitivity, and specificity (Table 3; Fig. 14). The best performance is achieved with 95% training data, where the accuracy reaches 0.915, sensitivity 0.865, and specificity 0.940. This demonstrates that a more extensive training dataset enhances the model’s predictive performance, making it a highly reliable approach for healthcare classification. Table 3Comparative Estimation concerning training Data.Training set %IterationAccuracySensitivitySpecificity60250.7150.7080.7660500.7350.7350.78560750.7650.7570.81601000.7950.7990.8470250.7270.720.7770500.7480.7480.80970750.8050.770.835701000.8520.810.88880250.7420.7350.7880500.7780.7650.8280750.8320.7880.85801000.8750.8250.990250.7590.7520.79590500.7990.780.83590750.8450.8050.875901000.8980.8480.9295250.7750.770.8195500.820.7980.8595750.870.830.89951000.9150.8650.94 Table 3Comparative Estimation concerning training Data.Training set %IterationAccuracySensitivitySpecificity60250.7150.7080.7660500.7350.7350.78560750.7650.7570.81601000.7950.7990.8470250.7270.720.7770500.7480.7480.80970750.8050.770.835701000.8520.810.88880250.7420.7350.7880500.7780.7650.8280750.8320.7880.85801000.8750.8250.990250.7590.7520.79590500.7990.780.83590750.8450.8050.875901000.8980.8480.9295250.7750.770.8195500.820.7980.8595750.870.830.89951000.9150.8650.94 Comparative Estimation concerning training Data. Fig. 14Comparative Estimation Concerning Training Dataset. Comparative Estimation Concerning Training Dataset. Comparative estimation concerning K value Table 3; Fig. 11 show that the TNBO + FCNN model consistently improves accuracy, sensitivity, and specificity as the K value increases. The best performance is observed with K = 11, where the accuracy reaches 0.905, sensitivity is 0.855, and specificity is 0.920. This suggests that a higher K value provides better model stability and generalization, making TNBO + FCNN a practical approach for healthcare classification tasks (Table 4; Fig. 15). Table 4Comparative Estimation concerning training data with various K-Fold.K-fold ValueIterationAccuracySensitivitySpecificity5250.750.720.775500.7780.7450.85750.810.7750.8351000.850.810.8657250.7650.7350.787500.790.7650.8157750.8320.7880.8571000.8750.8250.899250.780.750.7959500.8050.7780.8259750.8450.8050.8691000.890.840.90511250.7920.7650.80511500.820.790.8411750.860.820.88111000.9050.8550.92 Table 4Comparative Estimation concerning training data with various K-Fold.K-fold ValueIterationAccuracySensitivitySpecificity5250.750.720.775500.7780.7450.85750.810.7750.8351000.850.810.8657250.7650.7350.787500.790.7650.8157750.8320.7880.8571000.8750.8250.899250.780.750.7959500.8050.7780.8259750.8450.8050.8691000.890.840.90511250.7920.7650.80511500.820.790.8411750.860.820.88111000.9050.8550.92 Comparative Estimation concerning training data with various K-Fold. Fig. 15Comparative Estimation Concerning Training Data with Various K-Fold. Comparative Estimation Concerning Training Data with Various K-Fold. Comparative analysis To facilitate an equitable comparison, the efficacy of the proposed TNBO + FCNN model was assessed using established methodologies, such as Hybrid Multimedia Data Processing, RTS-DELM, Multimedia Data Sharing System, and SCA_WKNN. The benchmark models were either re-implemented under identical datasets and experimental conditions, or their outcomes were derived from existing scholarly publications when appropriate. For the models referenced in the literature, requisite modifications were made to ensure that the evaluation metrics aligned with our experimental framework. Although the comparison with existing models such as RTS-DELM and SCA_WKNN is clearly presented, the manuscript does not provide sufficient detail on the baseline configurations, dataset splits, or parameter tuning strategies used for these models. This lack of transparency may affect the perceived fairness and reproducibility of the comparisons. Including a table or subsection outlining the hyper parameters and training protocols for each baseline would enhance credibility and support stronger claims of superiority. Table 5; Fig. 16 compare the TNBO + FCNN method with existing approaches in terms of accuracy, precision, recall (sensitivity), F1-score, specificity, and AUC-ROC. Table 5Performance assessment.ModelAccuracyPrecisionRecall (Sensitivity)F1-ScoreSpecificityAUC-ROCHybrid Multimedia Data Processing0.7810.7540.7860.7700.8030.83RTS-DELM0.8370.8120.8250.8180.8280.85Multimedia Data Sharing System0.8520.8450.8400.8420.8510.87SCA_WKNN0.8720.8600.8550.8570.8740.88Proposed TNBO + FCNN0.9340.9320.9160.9230.9120.923 Table 5Performance assessment.ModelAccuracyPrecisionRecall (Sensitivity)F1-ScoreSpecificityAUC-ROCHybrid Multimedia Data Processing0.7810.7540.7860.7700.8030.83RTS-DELM0.8370.8120.8250.8180.8280.85Multimedia Data Sharing System0.8520.8450.8400.8420.8510.87SCA_WKNN0.8720.8600.8550.8570.8740.88Proposed TNBO + FCNN0.9340.9320.9160.9230.9120.923 Performance assessment. Fig. 16Performance Assessment of Different Models Vs Proposed Model. Performance Assessment of Different Models Vs Proposed Model. Training Set Size Effect: Larger training sets improve accuracy, sensitivity, and specificity, as they enable the model to learn more complex patterns with the additional data. K-Fold Effect: Increasing K improves model generalization by reducing bias, making the evaluation more robust. Comparison with Other Models: TNBO + FCNN outperforms existing models, achieving the highest accuracy (0.934), highest recall (0.916), and best specificity (0.932). F1-Score & AUC-ROC Performance: TNBO + FCNN achieves a balanced F1-score of 0.923 and a strong AUC-ROC of 0.923, confirming its superior classification performance. Thus, TNBO + FCNN is a highly effective and reliable model for healthcare classification tasks, offering greater accuracy, stability, and predictive power. While the TNBO + FCNN model demonstrates superior classification performance in accuracy, sensitivity, specificity, and AUC-ROC, inevitable trade-offs must be considered for real-world deployment. The model’s computational cost is higher than that of traditional methods due to the combined optimization of TNBO and deep feature extraction in FCNN. Additionally, inference speed is impacted by the complexity of processing multimodal data, making it less suitable for low-latency applications. However, hardware acceleration techniques such as GPU parallelization or model pruning can mitigate these computational demands. Future optimizations can focus on reducing complexity without compromising accuracy significantly. Discussion The proposed TNBO + FCNN model demonstrated significant improvements in accuracy, sensitivity, and specificity over traditional methods like RTS-DELM and SCA_WKNN, showcasing the effectiveness of hybrid deep learning in multimedia healthcare classification within IoT environments. Unlike previous approaches that struggled with scalability, our four-layer architecture enabled efficient and secure processing of large datasets. However, the limited diversity of the dataset is a constraint, and future studies should validate the model across broader demographics and conditions to enhance generalizability. Integrating block chain technology could further strengthen data security, particularly in IoT applications where privacy concerns are critical. We believe the TNBO + FCNN framework holds strong potential for real-time health monitoring in telemedicine platforms, with implications for policy development in healthcare data regulation. Despite the benefits of block chain-enabled IoT systems such as improved data integrity and patient privacy several ethical concerns persist: Data access and ownership ambiguity may lead to consent violations and unauthorized use.Risks of de-anonymization and inadequate anonymization across nodes challenge privacy.Cybersecurity threats remain due to vulnerable IoT devices and risks like 51% attacks.Opaque ML algorithms hinder informed clinical decisions and may reflect bias in healthcare data.Regulatory compliance is complex in decentralized networks, raising legal implications under frameworks like GDPR and HIPAA.Block chain immutability, while ensuring integrity, complicates the correction of erroneous or outdated records and raises concerns over long-term data stewardship. Data access and ownership ambiguity may lead to consent violations and unauthorized use. Data access and ownership ambiguity may lead to consent violations and unauthorized use. Risks of de-anonymization and inadequate anonymization across nodes challenge privacy. Risks of de-anonymization and inadequate anonymization across nodes challenge privacy. Cybersecurity threats remain due to vulnerable IoT devices and risks like 51% attacks. Cybersecurity threats remain due to vulnerable IoT devices and risks like 51% attacks. Opaque ML algorithms hinder informed clinical decisions and may reflect bias in healthcare data. Opaque ML algorithms hinder informed clinical decisions and may reflect bias in healthcare data. Regulatory compliance is complex in decentralized networks, raising legal implications under frameworks like GDPR and HIPAA. Regulatory compliance is complex in decentralized networks, raising legal implications under frameworks like GDPR and HIPAA. Block chain immutability, while ensuring integrity, complicates the correction of erroneous or outdated records and raises concerns over long-term data stewardship. Block chain immutability, while ensuring integrity, complicates the correction of erroneous or outdated records and raises concerns over long-term data stewardship. To ensure the proposed TNBO + FCNN framework can operate effectively in large-scale healthcare environments, several optimization strategies were considered. Model efficiency can be enhanced through parameter tuning, pruning, and quantization, enabling faster inference without significantly compromising accuracy. Batch processing and GPU-accelerated computations further reduce training time, while distributed learning architectures allow parallel processing across multiple nodes. Additionally, techniques such as incremental learning and feature caching can reduce retraining overhead when new patient data becomes available. These measures collectively enable deployment in real-time medical diagnostic systems, from high-performance cloud platforms to low-power IoT devices. Conclusion and future directions Conclusion The reliance on speech and images in healthcare classification poses limitations due to the narrow scope of available data and the diversity of medical conditions that require varied data types for accurate representation. This approach may lead to challenges in data quality, generalizability, and integration with other healthcare information. To enhance classification effectiveness, authors should broaden the dataset to include diverse medical data, such as electronic health records, diagnostic imaging, and wearable device inputs. A multimodal data strategy will enhance model robustness and accuracy, ultimately leading to improved patient outcomes. In the context of future scope, exploring potential advancements in security models for multimedia data sharing in IoT would be beneficial. This could involve further research into enhancing the robustness and efficiency of security measures, addressing emerging threats, and adapting to evolving IoT technologies. Additionally, considering the rapid development of IoT and multimedia technologies, future research could focus on integrating advanced encryption techniques, block chain technology, or artificial intelligence for more secure and efficient multimedia data sharing in IoT environments. Block chain provides a secure backbone for data management, enhancing trust in the model’s outputs by ensuring high-quality, unaltered data input. The proposed TNBO + FCNN model demonstrated significant improvements in healthcare data classification, achieving an accuracy of 92.4%, sensitivity of 92.1%, and specificity of 92.6%. These results surpass existing methods such as RTS-DELM and SCA_WKNN, confirming the model’s effectiveness in handling multimodal healthcare data. The high-performance metrics affirm the model’s potential for reliable real-time disease diagnosis and decision-making in IoT-enabled healthcare environments. Future work will focus on further optimizing the model’s “The combined TNBO and FCNN model’s inference complexity predominantly depends on the architecture of the FCNN, which, with optimized design, offers rapid predictions suitable for real-time healthcare monitoring. The inference operation’s complexity is proportional to the number of input features and neurons, generally resulting in low latency on modern hardware. Meanwhile, the TNBO optimization phase, which involves hyper-parameter tuning through iterative evaluations, is computationally intensive but occurs offline, thereby not impacting real-time deployment. This makes the model feasible for real-time systems, especially when deployed on edge or cloud servers, provided that the FCNN is designed with efficiency considerations in mind.” Limitations and future work While the proposed TNBO + FCNN framework demonstrates strong performance in classifying multimodal healthcare data, several limitations merit further consideration. First, the integration of TNBO introduces computational overhead, which may pose challenges for real-time or resource-constrained healthcare environments. Additionally, the blockchain layer, though beneficial for data integrity and security, can increase latency and processing complexity. The current evaluation is based on publicly available datasets, which may not fully capture the heterogeneity and noise present in real-world clinical data. Furthermore, the lack of clinical validation or field deployment limits the assessment of the model’s scalability, interoperability, and robustness in practical healthcare systems. Future work should focus on optimizing the model for real-time edge computing, evaluating cross-institutional generalizability, and incorporating federated learning techniques to enhance privacy while reducing transmission costs. Extensive clinical trials, longitudinal data testing, and hardware-aware implementations are also necessary to bridge the gap between theoretical performance and real-world applicability. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1