Quantum-enhanced federated blockchain for privacy-preserving cardiovascular intelligence Authors: R. Sivakami, V. Vinoth Kumar, N. Krishnamoorthy, Temesgen Engida Yimer Published: 2026 ABSTRACT Cardio-Vascular Diseases (CvDs) persist as a significant mortality reason worldwide, which requires advanced risk categorization technologies that can offer custom medicine while protecting patient information. Present healthcare approaches must overcome fragmented data distribution systems, weak security measures in collaborative environments, and the ineffective processing of multi-mode clinical details. Our proposed Decentralized Federated Blockchain for Cardiovascular Intelligence (DFBCI) framework integrates the Multi-Chain Aggregation with Adaptive Consensus (MCAC) mechanism and the advanced Federated Dynamic Relational Learning (FDRL) technique to solve identified challenges. The DFBCI system enables protected cooperation between institutions through its blockchain multi-chain setup that securely combines medical data with imaging information while maintaining data integrity. The FDRL technique extracts patient temporal behavior knowledge from different modalities, which include ElectroCardioGrams (ECGs), echocardiograms, and biomarker datasets. The system executes feature extraction using Quantum-Enhanced Privacy Masking (QEPM) for security alongside non-invasive cardiac risk modeling through federated optimization and performs decentralized validation with smart contracts. Compared to standard approaches, Experimental results reveal that models achieve 19% better CvD risk prediction while reducing their convergence time by 22% and enhancing scalability by 27%. DFBCI creates powerful healthcare analytical platforms that advance secure CvD risk sorting worldwide without violating privacy rights, setting benchmark for fast, collaborative assessments. Introduction Cardiovascular Diseases (CvD) stand as the major global death-cause requiring modern AI-developed protected risk prediction systems1. Healthcare digitalization enables machine learning (ML) along with deep learning (DL) tools for medical analytics that lead to diagnostic and prognostic transformations. Federated Learning (FL) and blockchain (Bc)-based decentralized intelligence systems attract attention because they allow joint analytics in healthcare without releasing important patient data sets2. Quantum cryptography serves as one of the multiple developing privacy-protecting methods that protect patient information from cyber threats3. The unified system of FL, Bc and DL technologies provides healthcare facilities with secure prediction accuracy capabilities4. The field of healthcare AI currently focuses on data protection through multiple significant approaches. Healthcare institutions implement federated learning as their technique for training AI models at different sites while maintaining data privacy through location-based control. The implementation of blockchain technology serves to enable safe exchanges of health data between different parties5. Scientists have developed quantum encryption strategies that aim to provide better privacy protection for confidential medical information6. The present solutions emerge with multiple problems involving data disintegration while having poor results regarding privacy techniques and combined clinical information processing. AI-driven CvD risk prediction remains limited due to insufficient decentralized frameworks that ensure a secure combination of electrocardiogram (ECG) signals and echocardiogram images and biomarker datasets7. The novelty of this study stems from Decentralized Federated Blockchain for Cardiovascular Intelligence (DFBCI) to span the five independent operational phases. Blocks of Multiple blockchains together with Adaptive Consensus mechanism function as Multi-Chain Aggregation with Adaptive Consensus (MCAC) in blockchain-powered secure data transmission while Quantum-Enhanced Privacy Masking (QEPM) enables privacy preservation and Federated Dynamic Relational Learning (FDRL) performs multi-modal feature analysis and multi-variable deep learning allows predictive modeling. Determining the validity of AI forecasting models through blockchain-integrated smart contracts guarantees tamper-evidence prediction outcomes. The innovative system demonstrates better performance compared to conventional AI techniques in CvD risk evaluation, ensures heightened security for patient information, and handles extensive datasets. Significance and contribution The DFBCI framework stands vital within healthcare applications when implementing AI to predict cardiac vascular diseases. The traditional predictive models store and process data in centralized locations, creating challenges for patients’ information protection, institutional communication barriers, and facility security8. DFBCI enables blockchain-based federated learning to empower multiple healthcare institutions in AI model collaboration without disclosing patient information at its core level. The framework meets protection regulations requirements; thus, it provides real-world clinical applications for deployment. The DFBCI framework contributes significant social effects through its privacy-preserving methods for deploying AI-driven CvD risk predictions. The combination of federated learning with blockchain technology in DFBCI creates safe, standard medical AI applications that support cross-institutional healthcare collaboration and patient privacy protection9. DFBCI plays an essential role in society by establishing a platform for health data exchange between facilities that serve disadvantaged and poorly equipped communities. Developing areas cannot utilize superior AI-based CvD risk prediction tools because they struggle with both privacy-related challenges and weak infrastructure capabilities10. People conducting epidemiological studies and public health studies can access encrypted cardiovascular patient data through DFBCI without damaging privacy standards. The DFBCI system delivers superior CvD risk prediction performance while decreasing runtime and raising its operational capacity to achieve standards beyond existing privacy-protecting AI healthcare systems. Real-time disease trend analysis supported by this method enables policymakers to make evidence-based decisions and create CvD prevention methods at the global level11. Smart contract-based model validation through tamper-proof and unbiased mechanisms helps eliminate wrong clinical choices that result from adversarial model tampering. Scope and motivation The investigation scope covers present clinical choice procedures alongside healthcare data combination methods from multiple facilities and AI applications for cardiovascular risk estimation. The worldwide highest mortality rate belongs to cardiovascular diseases, which require immediate early detection and risk assessment to deliver timely intervention and treatment12. Present CvD risk models that use AI do not manage patient data across different sources unless hybridized effectively, nor do they operate with secure methods, so they prove ineffective for medical use. The DFBCI system solves these challenges through its approach that provides flexible deployment of privacy-focused distributed AI intelligence within clinical settings, research facilities, and remote healthcare centers. The rising healthcare requirement for privacy-protecting AI creates the motivation to develop methods that enable institutions to utilize sophisticated AI models while maintaining patient privacy. The current surge of digital attacks, together with non-trustworthy AI implementations and unprotected data, shows why healthcare institutions need secure AI deployment procedures for maintaining responsible data management13. DFBCI framework emerges from a combined ideology between federated learning and blockchain and quantum cryptography, which introduces a hybrid approach toward secure medical analytic systems using AI. The research presents methods to combine intelligence with data protection through a framework that enables sustainable healthcare progress that follows ethical AI frameworks and medical needs. Objectives. To design DFBCI as a blockchain-based federated learning framework that provides protected and distributed CvD risk evaluation and confirms data authenticity and user privacy.To enhance model performance by utilizing FDRL and multi-variable deep neural networks and to improve both feature extraction and predictive accuracy.To secure AI forecasting among medical collaborators through a decentralized model validation mechanism based on smart contracts (SC) to shield from unauthorized changes. To design DFBCI as a blockchain-based federated learning framework that provides protected and distributed CvD risk evaluation and confirms data authenticity and user privacy. To design DFBCI as a blockchain-based federated learning framework that provides protected and distributed CvD risk evaluation and confirms data authenticity and user privacy. To enhance model performance by utilizing FDRL and multi-variable deep neural networks and to improve both feature extraction and predictive accuracy. To enhance model performance by utilizing FDRL and multi-variable deep neural networks and to improve both feature extraction and predictive accuracy. To secure AI forecasting among medical collaborators through a decentralized model validation mechanism based on smart contracts (SC) to shield from unauthorized changes. To secure AI forecasting among medical collaborators through a decentralized model validation mechanism based on smart contracts (SC) to shield from unauthorized changes. Related work The author in14 introduced a secure healthcare 5.0 system that combines FL with IDS security to protect privacy while using Blockchain for real-time disease prediction. Real-Time Deep Extreme Learning Machine (RTS-DELM) provides the methodology with an efficient tool for detecting intrusions and predicting diseases in healthcare networks. The research used Parkinson’s disease and the NSL-KDD dataset to reach 93.22% disease prediction accuracy together with 96.18% intrusion detection accuracy. The system connects different components by using blockchain technology for network security, together with federated learning protocols for distributed model creation and intrusion detection systems for threat monitoring. This system provides privacy protection for medical patient monitoring devices in the IoMT framework through diagnostic accuracy enhancement and better data security. The proposed method suffers from three main drawbacks, including elevated processing time requirements and additional blockchain-related operational burden, and possible attacks against model parameters. Researchers in15 introduced FedEHR, which applied FL techniques to heart disease predictions within electronic health records (EHRs) from IoT devices by employing Cluster Primal-Dual Splitting (cPDS) together with soft-margin L1-regularized Support Vector Machine (sSVM) as an optimization method to optimize efficiency and computational performance. The research deploys FL to deliver distributed predictive functionalities across multiple systems without breaking data confidentiality rules. The research analyzes heart disease predictions by uniting IoT health data with EHRs through FL methods, which prove more accurate and efficient than conventional centralized analysis while reaching a peak accuracy rate of 99.86% through 1000 communication cycles, better than GA-SVM, Fed-SVM, and SVM baseline systems. Through its operation, the cPDS algorithm fends off excessive data transfer demands while sustaining excellent classifying outcomes. The implementation of this approach comes with several constraints that involve considerable start-up expenses, delayed model maintenance, and difficulties when adapting the model to handle big healthcare systems. The work in16 presented TrustDFL, which serves as an FL framework that utilizes blockchain to create verifiable decentralized FL systems through zk-SNARKs and smart contracts with IPFS functionality. The methodology uses TrustDFL to incorporate blockchain as a trust foundation so users can perform verifiable privacy-protected model training without exposing their data sources. The investigation aims to boost FL reliability through attack prevention and fair model update verification using cryptographically secure methods. The experimental outcomes verify that TrustDFL enables accurate global outcomes by stopping harmful model modifications at a minimal computational verification expense. The system decreases on-chain storage demands by keeping only hash values in place of large model files that exist on IPFS. The zk-SNARK verification method, together with blockchain-induced latency and proof generation time, increases while also generating substantial computational expenses. Authors17 introduced an advanced healthcare 5.0 security platform by uniting Matrix-Valued Neural Coordinated Federated Intelligence with Blockchain and Intrusion Detection Systems (IDS) for disease prediction and Internet of Healthcare Things (IoHT) intrusion detection purposes. Federated Deep Extreme Machine Learning (FDEML) serves as the methodology where distributed training takes place to safeguard privacy across all participating hospitals. The proposed methodology succeeds in disease forecasting with a 98% accuracy rate by utilizing the Parkinson’s dataset along with the NSL-KDD dataset for network security threat detection, which reaches 97.75% efficiency. The Blockchain security layer protects model update safety while maintaining data consistency and enabling transparent financial operations. IDS operates as an always vigilant threat detection system that stops cyber-attacks. The Neural Synchronization method allows users to generate secure cryptographic keys for defense against conventional key exchange system weaknesses. This approach deals with high computational demands and blockchain implementation issues, and attacks on model updates and neural key synchronization (NKS) performance problems. Researchers18 proposed a data privacy solution for IoMT applications based on Blockchain technology combined with RL and FL to develop a secure system for clinical organization monitoring. The method utilizes blockchain-enabled reinforcement FL to protect patient data safely and process it on a large scale with minimal server dependency. Distributed multi-disciplinary reinforcement FL analyzes post-COVID-19 patient data with excellent performance while surpassing all current privacy-protected methodologies. Blockchain transactions remain immutable through its integration, and reinforcement learning (RL) supports immediate decision-making processes for IoMT platforms. System performance shows better results along with improved security features, which makes the system fit for healthcare environments requiring privacy protection. The method has several issues that involve significant computational needs, Blockchain scalability problems, and potential slowing down during RL convergence. Investigators19 designed a secure federated cloud storage system for IoMT applications through Hybrid Heuristic Attribute-Based Encryption (HHABE) and Permissioned Blockchain, which strengthened data security privacy and regulatory compliance. The system generates optimal ABE keys by implementing Hybrid Mexican Axolotl with Energy Valley Optimizer (HMO-EVO) as its key generation method, which results in both fast encryption and low additional processing requirements. This work targets two goals: first, it protects medical records, and then it makes possible federated disease prediction through Multi-Scale hybridized recurrent models that operate on diabetes and heart disease datasets. The new system delivers superior prediction accuracy, improved security, shorter storage times, and encryption time compared to existing models. The outcome exhibits faster cryptographic operations and low computational overhead with complete security against data thefts. Implementing this method is challenging due to its expensive setup requirements, blockchain-related delay issues, and computational complexity during extensive deployments. The study in20 developed EdgeGuard as a distributed healthcare resource control system that connects FL algorithms with Blockchain technology and Edge Computing to achieve secure Internet of Medical Things operations and extend scalability. The methodology implements a privacy-protected FL algorithm with differential privacy enhancement that uses adaptive aggregation protocols alongside a lightweight blockchain protocol that secures medical data. This solution framework establishes methods to arrange medical devices, protect model analysis performance throughout IoMT networks that span various geographic areas, and implement RL edge resource distribution protocols and multi-party computational mechanisms for model sharing. Real-world medical datasets reveal that EdgeGuard reaches 94.3% model accuracy together with 30.67% communication overhead reduction and system security protection against up to 40.05% malicious nodes. This approach demonstrates restricted usage because it necessitates extensive calculations, raises blockchain-related energy usage, and creates difficulties in balancing privacy protection with performance capabilities. The privacy-enhanced heart disease prediction framework presented in21 combines FL with Blockchain technology through the application of TabNet computational mode, which unites deep learning operations with decision tree approaches to improve interpretability while maintaining predictive accuracy. The research design implements FedAvg to perform horizontal federated learning across institutions, which protects patient privacy through decentralized training procedures. The research uses the Comprehensive Heart Disease and UCI Heart Disease datasets alongside TabNet to perform heart disease predictions through its enabled feature selection and representation learning capabilities. TabNet reaches 82.2% accuracy, accompanied by a 1.594 balanced metric score after 50 network training cycles, and maintains a robust performance during 10 aggregation cycles of integrated data streams. Through smart contracts applied by blockchain technology, the system achieves enhanced security and transparency that ensures model updates are protected. The method has two main drawbacks, including expensive computations, data heterogeneity challenges, and the time needed for blockchain management operations. The work in22 developed a privacy-protected framework that combines FL with blockchain technology for medical chest disease classification to tackle essential medical data security and privacy and diagnostic accuracy needs. FL serves as the methodology because it enables multiple healthcare institutions to jointly train models through decentralized learning while preserving individual patient data privacy during non-disclosure operations. Privacy protection and tamper prevention are achieved through blockchain integration, which also enables secure model update transparency and data transactions. This framework employs deep learning models optimized through federated learning for the multi-classification of chest diseases that include COVID-19, pneumonia, and lung cancer. It reaches classification accuracy levels from 91.26% to 99.51% with precision at 91.40% to 97.70% and up to 97.99% F1-score. The method delivers better performance than centralized deep learning systems, with major advantages in privacy preservation that make it applicable to clinical practices that need strong data protection. Broader FL implementation requires solutions to interoperability problems and optimization of computational expenses and regulatory compliance challenges. The work in23 established Distributed Cross-Learning for Equitable Federated Models (D-CLEF) as a decentralized framework that unites FL and Blockchain to protect privacy during equitable healthcare predictive analysis. D-CLEF serves as the methodology that supports data partitions across both horizontal and vertical dimensions to facilitate medical institution collaboration with protected patient records. The study demonstrates how D-CLEF performs similarly to centralized learning when predicting COVID-19 mortality alongside surgical outcomes and myocardial infarction by analyzing five California hospitals and one Edinburgh heart disease dataset and proves better outcomes than independent isolated models. D-CLEF matches centralized learning performance while adding a minimal 10% delay during synchronization, which leads to effective joint work between healthcare institutions across horizontal and vertical connections. The implementation of blockchain in the system creates transparency and scalability and ensures that it models everything from a single source while eliminating control vulnerabilities. The approach encounters challenges related to elevated synchronization times along with blockchain-caused computational burden, and difficulties when applying D-CLEF to multiple healthcare formats, including medical images. The different approaches and outcomes are listed in below Table 1. Table 1Summarized overview of existing approaches.ReferencesApproachComponentsFocusedOutcomesLimitations 14 RTS-DELM + FLBcFL, Blockchain, IDS, IoMT SensorsDisease prediction and intrusion detectionDisease Prediction Accuracy: 93.22%, Intrusion Detection Accuracy: 96.18%vulnerability to adversarial attacks, blockchain-induced processing overhead 15 FedEHRFL, EHRs, sSVM, cPDSHeart Disease Prediction in IoT-Based EHRsHigh accuracy: 99.86% after 1000 communication roundsScalability challenges, Model synchronization delays, and High setup costs in large-scale healthcare systems 16 TrustDFLFL, Blockchain, zk-SNARKs, SC, IPFSSecuring FL against model poisoning attacks and ensuring verifiable model trainingMaintains model accuracy, reduces storage burden using IPFS, prevents malicious updatesBlockchain latency, increased proof generation time, high computational cost of zk-SNARK verification 17 FDEML + BcFL, Blockchain, NKS, IDS, IoHTDisease Prediction, Intrusion DetectionDisease Prediction Accuracy: 98%, Detection Efficiency (Intrusion): 97.75%Blockchain overhead, adversarial attacks, NKS inefficiency 18 RFLBlockchain, FL, RL, Decentralized Privacy MechanismsPost-COVID-19 Patient Data Monitoring in IoMTHigh reliability, improved security, better scalability than existing modelsBlockchain scalability issues, RL inefficiencies 19 HHABE + BcFL, Blockchain, HMO-EVO, Multi-Scale recurrent modelsDisease Prediction (Diabetes, Heart Disease), Secure Federated Cloud Storage for IoMTHigh prediction accuracy, improved encryption efficiency reduced storage overheadHigh initial setup cost, blockchain latency, computational overhead for large-scale deployments​ 20 EdgeGuardAdaptive Aggregation, Differential Privacy, RL, Lightweight ConsensusMedical Resource Orchestration & Secure Model Training in IoMT NetworksAccuracy: 94.3%, Security Against Malicious Nodes: 40.05%, Communication Overhead Reduction: 30.67%High computational complexity, blockchain-induced energy consumption, trade-offs between privacy and performance​ 21 TabNet + FLBcFeature selection via attention mechanism, FedAvg, Differential Privacy, Smart ContractsHeart Disease PredictionAccuracy: 82.2%, Balanced Metric: 1.594Data heterogeneity, blockchain-induced training delays​ 22 FLBcDecentralized training, privacy preservation, data securityClassification of chest diseasesAccuracy: 91.26% − 99.51%, Precision: 91.40% − 97.70%, F1-Score: 96.65% − 97.99%High computational cost, regulatory compliance, interoperability challenges​ 23 D-CLEFFL, Blockchain, Horizontal & Vertical Data Partitioning, Decentralized model trainingCOVID-19 Mortality Prediction, Myocardial Infarction, Prediction Surgical OutcomesPerformance equivalent to centralized learning, 10% increase in synchronization time, outperforming existing modelsIncreased synchronization time, blockchain computational overhead, challenges in integrating multi-modal healthcare data​ Table 1Summarized overview of existing approaches.ReferencesApproachComponentsFocusedOutcomesLimitations 14 RTS-DELM + FLBcFL, Blockchain, IDS, IoMT SensorsDisease prediction and intrusion detectionDisease Prediction Accuracy: 93.22%, Intrusion Detection Accuracy: 96.18%vulnerability to adversarial attacks, blockchain-induced processing overhead 15 FedEHRFL, EHRs, sSVM, cPDSHeart Disease Prediction in IoT-Based EHRsHigh accuracy: 99.86% after 1000 communication roundsScalability challenges, Model synchronization delays, and High setup costs in large-scale healthcare systems 16 TrustDFLFL, Blockchain, zk-SNARKs, SC, IPFSSecuring FL against model poisoning attacks and ensuring verifiable model trainingMaintains model accuracy, reduces storage burden using IPFS, prevents malicious updatesBlockchain latency, increased proof generation time, high computational cost of zk-SNARK verification 17 FDEML + BcFL, Blockchain, NKS, IDS, IoHTDisease Prediction, Intrusion DetectionDisease Prediction Accuracy: 98%, Detection Efficiency (Intrusion): 97.75%Blockchain overhead, adversarial attacks, NKS inefficiency 18 RFLBlockchain, FL, RL, Decentralized Privacy MechanismsPost-COVID-19 Patient Data Monitoring in IoMTHigh reliability, improved security, better scalability than existing modelsBlockchain scalability issues, RL inefficiencies 19 HHABE + BcFL, Blockchain, HMO-EVO, Multi-Scale recurrent modelsDisease Prediction (Diabetes, Heart Disease), Secure Federated Cloud Storage for IoMTHigh prediction accuracy, improved encryption efficiency reduced storage overheadHigh initial setup cost, blockchain latency, computational overhead for large-scale deployments​ 20 EdgeGuardAdaptive Aggregation, Differential Privacy, RL, Lightweight ConsensusMedical Resource Orchestration & Secure Model Training in IoMT NetworksAccuracy: 94.3%, Security Against Malicious Nodes: 40.05%, Communication Overhead Reduction: 30.67%High computational complexity, blockchain-induced energy consumption, trade-offs between privacy and performance​ 21 TabNet + FLBcFeature selection via attention mechanism, FedAvg, Differential Privacy, Smart ContractsHeart Disease PredictionAccuracy: 82.2%, Balanced Metric: 1.594Data heterogeneity, blockchain-induced training delays​ 22 FLBcDecentralized training, privacy preservation, data securityClassification of chest diseasesAccuracy: 91.26% − 99.51%, Precision: 91.40% − 97.70%, F1-Score: 96.65% − 97.99%High computational cost, regulatory compliance, interoperability challenges​ 23 D-CLEFFL, Blockchain, Horizontal & Vertical Data Partitioning, Decentralized model trainingCOVID-19 Mortality Prediction, Myocardial Infarction, Prediction Surgical OutcomesPerformance equivalent to centralized learning, 10% increase in synchronization time, outperforming existing modelsIncreased synchronization time, blockchain computational overhead, challenges in integrating multi-modal healthcare data​ Summarized overview of existing approaches. Methodology DFBCI architecture in Fig. 1 exhibits the usage of the Multi-Chain Aggregation with Adaptive Consensus (MCAC) technique for distributed secure CvD data storage that conducts real-time assessments and supports multisource data combinations while protecting data from tampering. QEPM provides privacy protection to critical medical information while maintaining accessible non-invasive cardiac risk evaluations for patients. The FDRL module extracts time patterns from ECGs and echocardiograms and biomarkers through protected educational learning methods. Through its execution process, Decentralized Smart Contract Execution performs automated model validation alongside trustless auditing functions. The cloud-edge system architecture can perform real-time analysis for efficient yet compliant and scalable CvD risk assessments. Fig. 1Architecture of DFBCI. Architecture of DFBCI. Figure 2 diagrammatically exhibits the operational workflow of the Decentralized Federated Blockchain for Cardiovascular Intelligence (DFBCI) framework with all vital components of various phases that poses significant influences in the attainment of objective task. DFBCI stands as a new methodological framework that combines blockchain-powered FL with quantum encryption protocols and multiple DL variables for better CvD risk forecasting. Fig. 2Operational workflow of DFBCI framework across six vital phases. Operational workflow of DFBCI framework across six vital phases. The Multi-Chain Aggregation with Adaptive Consensus (MCAC) mechanism establishes blockchain-based, secure, decentralized data-sharing by permitting multiple healthcare institutions to verify patient records within blockchain-powered tamper-proof systems. Each institution maintains its own blockchain sub-chain under MCAC, which safeguards secure data exchange between institutions, and the adaptive consensus algorithms assess credibility when validating data sources. Quantum-Enhanced Privacy Masking (QEPM) protects medical records by using QKD encryption together with QNA to add noise and to de-identify delicate information that must remain accessible for risk evaluation processes. After securing the data into the system, the Federated Dynamic Relational Learning (FDRL) component detects temporal patterns and intricate connections between different cardiovascular datasets that include ECG time-series, echocardiogram sequences, and biomarker changes. Each institution can maintain data ownership during model development through this decentralized approach because they control the resources for AI model evolution toward better feature acquisition without requiring data consolidation. A deep neural network (DNN) functions at the Multi-Variable Integration for the CvD Risk Modeling phase to combine several cardiovascular parameters into a unified predictive structure. The risk stratification process receives optimized benefits through federated optimization technology that enables dynamic multi-modal learning, which promotes early disease detection. The Decentralized Model Validation using Smart Contracts implements blockchain components for secure automatic verification of AI forecasting through embedded model testing systems24. The smart contracts of this system conduct automatic model update validation through specified accuracy and convergence threshold requirements, which stop unauthorized interference and allow doctors to receive trusted AI recommendation services. The multi-tier architecture of DFBCI makes significant progress in accurate CvD risk assessment both during real-time clinical decisions and enhances protected collaborative intelligence and attainable scalability, thereby positioning itself as a benchmark tool for secure AI assessments for contemporary healthcare. The following section provides a brief, precise explanation of five core phases of the framework with all essential computations. They are. Phase-1: Data Acquisition and Secure Multi-Chain Aggregation (DA-SMA).Phase-2: Quantum-Enhanced Privacy Masking (QEPM).Phase-3: Federated Dynamic Relational Learning (FDRL).Phase-4: Multi-Variable Integration (MVI) for CvD Risk Modeling.Phase-5: Decentralized Model Validation using Smart Contracts (DMV-SC). Phase-1: Data Acquisition and Secure Multi-Chain Aggregation (DA-SMA). Phase-1: Data Acquisition and Secure Multi-Chain Aggregation (DA-SMA). Phase-2: Quantum-Enhanced Privacy Masking (QEPM). Phase-2: Quantum-Enhanced Privacy Masking (QEPM). Phase-3: Federated Dynamic Relational Learning (FDRL). Phase-3: Federated Dynamic Relational Learning (FDRL). Phase-4: Multi-Variable Integration (MVI) for CvD Risk Modeling. Phase-4: Multi-Variable Integration (MVI) for CvD Risk Modeling. Phase-5: Decentralized Model Validation using Smart Contracts (DMV-SC). Phase-5: Decentralized Model Validation using Smart Contracts (DMV-SC). Phase-1: DA-SMA The initial stage of DFBCI collects multi-component cardiovascular data through healthcare institutions which convert data into electrocardiograms (ECGs)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({I}_{ECG}\in\:{\mathrm{R}}^{\left(n\times\:{T}_{ECG}\right)}\right)$$\end{document}, echocardiograms \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({I}_{echo}\in\:{\mathrm{R}}^{\left(n\times\:{T}_{echo}\right)}\right)$$\end{document}and biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({I}_{bio}\in\:{\mathrm{R}}^{\left(n\times\:m\right)}\right)$$\end{document} datasets with respect to the time dimension (T), number of patients (n) and biomarkers (m). Healthcare data distributed across multiple sources needs aggregation through the Multi-Chain Aggregation with Adaptive Consensus (MCAC) solution to achieve secure data combination. Every institution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\mathbb{I}}_{i}\right)$$\end{document} runs its unique blockchain sub-chain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\beta\:}_{i}\right)$$\end{document} to securely (encoding) store indelible medical transaction data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\mathrm{T}}_{ij}\right)$$\end{document} for any patient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\rho\:}_{j}\right)$$\end{document}, which is expressed as,1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbb{B}}_{i}=\left\{\left[{\mathrm{h}}\left({\mathrm{T}}_{i1}\right),{\mathrm{h}}\left({\mathrm{T}}_{i2}\right),\cdots\:,{\mathrm{h}}\left({\mathrm{T}}_{ik}\right)\right]\right\}$$\end{document} Database aggregation through adaptive consensus (C) guarantees data integrity by validating source institutions before data collection procedure begins.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:C=\mathrm{arg}\underset{C}{\mathrm{max}}{\sum\:}_{i=1}^{N}\left[\varPsi\:\left({\beta\:}_{i}\right)\times\:{\omega\:}_{i}\right]$$\end{document} From (2), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\omega\:}_{i}$$\end{document} denotes the credibility weightage of ith institution and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varPsi\:\left({\beta\:}_{i}\right)$$\end{document} indicates the measure of utility function in determining the trustworthiness. This phase resolves data interoperability problems by enabling safe institution-to-institution information sharing across decentralized networks which prevent tampering of data. Phase-2: QEPM Doctor-patient privacy, together with data security, stands as a vital issue when managing cardiovascular health information. The healthcare system uses QEPM to protect patient confidentiality by using quantum cryptographic procedures for data encryption before model training. QKD establishes secure encryption for medical transactions by protecting against unauthorized access to private information. This research study implements QKD encryption using BB84 protocol across each institution with procedures defined as25,3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|{\mathrm{q}}\rangle \right.=\mu\:\left|0\rangle \right.+\phi\:\left|1\rangle \right.,\:\:\:\:where\:\:{\left|\mu\:\right|}^{2}+{\left|\phi\:\right|}^{2}=1$$\end{document} The BB84 protocol operates as a QKD protocol that allows secure encryption. From (3), the quantum state functions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\left|{\mathrm{q}}\rangle \right.\right)$$\end{document} as a superposition between basis states \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|0\rangle \right.$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|1\rangle \right.$$\end{document} with probability amplitudes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\pi\:$$\end{document}, which add up to one. Quantum mechanics enables secure key exchange through its mechanism and provides resistance against eavesdropping. The random rotation of qubits occurs through Hadamard transformation26, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}^{\tau\:}$$\end{document} and is described as in Eqs. (4),4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}^{\tau\:}={\left(\sqrt{2}\right)}^{-1}\cdot \:\left[\begin{array}{cc}1&\:1\\\:1&\:-1\end{array}\right]$$\end{document} Quantum noise addition (QNA) is deployed to conceal patient data which enables the preservation of necessary features for ML processing methods. Gaussian noise function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{g}}\sim\mathrm{N}\left(0,{\sigma\:}^{2}\right)$$\end{document}27 is employed to serves as the privacy protection method during such data masking operations and is expressed as,5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{masked}=I+{\mathrm{g}}$$\end{document} The Eq. (5) demonstrates the privacy masking progress in adding noise to data in order to achieve information protection. Secure features in the data remain usable after masking it enables protection of sensitive information. The method protects secure computations and federated learning procedures because it blocks unauthorized individuals from accessing raw data formats. Thus, the model performance remains unaffected by implementing this step which establishes data protection from breaches and unauthorized access in federated environments. Phase-3: FDRL The FDRL technique extracts temporal relationships and complex data dependencies between modalities after the completion of privacy masking. DL architectures analyze cardiovascular risk factors through time because these factors change with time. Healthcare institutions at the local level obtain multi-modal patient information that includes ECG signals together with echocardiograms and biomarker measurement values. Each institution maintains central processing of its raw data through deep learning models instead of sending it to a centralized server. In this study, Temporal Convolutional Networks (TCNs) operate on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{ECG}$$\end{document} time-series to discover systemic patterns that enable the detection of abnormal heart rhythm irregularities28. The core significant operational process of TCN in hidden layers is expressed as,6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}_{t}=A\left[\left({I}_{t}\cdot \:\omega\:\right)+e\right]$$\end{document} The 3D Convolutional Neural Networks (3D CNNs) process echocardiogram frames \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{echo}$$\end{document} to find heart structural details, which help detect cardiovascular problems29. The primary feature extraction computation is expressed as,7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{echo}=ReLU\left[\left({I}_{echo}\cdot \:\omega\:\right)+e\right]$$\end{document} Institutional servers maintain local training of models \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\Gamma_{i}^{t}$$\end{document} that protect patient privacy by allowing customized model development using native hospital database information.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\Gamma_{i}^{t}=-\gamma\:\nabla\:\mathrm{L}\left(\Gamma_{i}^{t-1};{I}_{i}\right)+\Gamma_{i}^{t-1}$$\end{document} The global system aggregates knowledge from local models without revealing patient data. Through the Federated Averaging (FedAvg) method30, every institution sends its model updates to a central global server, which is computed as,9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Gamma}^{t}={\sum\:}_{i=1}^{N}\left(\raisebox{1ex}{${d}_{i}$}\!\left/\:\!\raisebox{-1ex}{$N$}\right.\right)\cdot \:\Gamma_{i}^{t}$$\end{document} From (9), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}$$\end{document} indicates the data samples from ith institution and N represents the total count of samples across all institutions. The server operates solely with model updates from local facilities yet avoids processing actual patient data. Several institutions send their updates to create a superior global model utilizing aggregate intelligence from different organizations. A distributed framework to handle patient information achieves better prediction accuracy for CvD detection apart from safeguarding privacy, which enables healthcare monitoring in diverse populations. The global model update travels back to each institution where clinicians can access recent advancements without revealing their protected health information. Phase-4: MVI for CvD risk modeling A multi-variable integration is employed via a deep neural network (DNN) to attain an accurate CvD risk prediction. The multi-variable integral form of risk function31, r(CvD) is defined as follows,10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:r\left(CvD\right)={\int\:}_{\mathrm{Z}}f\left({I}_{ECG},{I}_{echo,\:}{I}_{bio}\right)dX$$\end{document} On applying triple integration decomposition to Eqs. (10),11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:r\left(CvD\right)={\int\:}_{{t}_{1}}^{{t}_{2}}{\int\:}_{{f}_{1}}^{{f}_{2}}{\int\:}_{{b}_{1}}^{{b}_{2}}f\left({I}_{ECG},{I}_{echo,\:}{I}_{bio}\right)dbdfdt$$\end{document} From (11), the domain specification Z includes all potential input variable values for which integration takes place. Time is symbolized by t to indicate the duration of ECG signal modifications. The risk function f establishes relationships between ECG and echocardiogram and biomarker data to predict CvDs. The term b represents biomarker values that serve as vital markers to evaluate cardiovascular disease risk levels. During this stage, the model unites various health indicators through a single prediction framework by processing ECG signals together with imaging data and biochemical markers. This DNN training process occurs on federated datasets to develop predictions that both maintain accuracy levels and generalize across all institutions. Special optimization algorithms minimize bias and develop robust learning capabilities, and the model presents adaptive features for handling changes in individual patient data. The combination of various health-related data types in one predictive model system enhances risk assessment abilities and raises detection timeframes for better early cardiovascular disease intervention. Table 2 represents the DNN training and prediction processes in an algorithmic way. Table 2Core algorithmic procedures of DNN training and prediction. Table 2Core algorithmic procedures of DNN training and prediction. Core algorithmic procedures of DNN training and prediction. Phase-5: decentralized model validation Model integrity protection, along with the prevention of malicious changes, is achieved through blockchain-based smart contracts (SC) that operate as decentralized validators within the DFBCI framework. The SC performs decentralized validation through its protocol, which checks the relative changes between updated models. The SC accepts parameter updates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Gamma}^{t}$$\end{document} when the normalization factor of current to previous parameters shows a change below a safety threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varepsilon$$\end{document} and rejects other changes. The decentralized learning system only receives meaningful and stable model improvements after validation procedures. Thus, the core validation process in the SC is exhibited as,12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:SC\left({\Gamma}^{t}\right)=\left\{\begin{array}{c}accept,\:\:if\frac{\|{\Gamma}^{t}-{\Gamma}^{t-1}\|}{\|{\Gamma}^{t-1}\|}<\epsilon\:\\\:reject,\:\:otherwise\end{array}\right.$$\end{document} From (12), it is evident that the SC independently verify predictive model reliability and consistency before deployment takes place. The system ensures secure model updates by implementing contracts that test fulfillment of specified convergence criteria to block unauthorized or ineffective updates before deployment. This final phase maintains transparentness, secureness, and verifiability, thus building trust in CvD risk assessments. Table 3 highlights the core operational processes of the SC in an executable code snippets. Table 3Executable code snippets of SC process. Table 3Executable code snippets of SC process. Executable code snippets of SC process. Material utilized The DFBCI framework requires specific hardware specifications and software elements to implement FL alongside quantum-enhanced privacy masking and blockchain operations, and DL-based CvD risk prediction. The testbed platform implements Python v3.9 + as its main coding language, while PyTorch v1.10 + serves as the framework for developing DL models. The decentralized training needs Flower v1.2+ (FL strategies), whereas secure data transactions require Hyperledger Fabric v2.5 + blockchain platforms. The QEPM module integrates Qiskit v0.39 + as an open-source quantum computing library for QKD and QNA functionalities. The implementation of model validation and blockchain-embedded SCs requires Python with the Solidity v0.8+. Data preprocessing, together with analytics, runs on Pandas v1.4 + while NumPy v1.23 + and SciPy v1.9+, along with Matplotlib v3.5 + and Seaborn v0.12+, support the visualization processes. The study executed its operations on a high-performance GPU-enabled system that uses an NVIDIA A100 as the minimum specification to achieve efficient DL task execution. The system selects the multi-core CPU AMD-EPYC 7543-P as the baseline processor for federated training operations. Quick data processing depends on NVMe SSD storage devices that have at least 2 TB storage capacity together with 128GB RAM to manage extensive medical databases effectively. The blockchain network depends on Docker 24 + and Kubernetes 1.24 + to run on Ubuntu 20.04 LTS Linux servers through their cluster in cloud infrastructure. The proposed system provides an efficient CvD risk prediction solution with scalability and privacy and meets all requirements for open-source frameworks to ensure transparency and reproducibility. Table 4 exhibits the hyperparameters for all components in the DFBCI framework. Demonstrating hyperparameters provides essential benefits for empirical evaluation since it enables reproducibility and model optimization and performs benchmarking across different configurations. Proper optimization of components allows researchers to validate the execution parts of AI models and FL and blockchain elements in the framework. Table 4Configuration of DFBCI framework’s hyperparameters.ComponentsHyperparameterOptimal valueFDRLLearning rate0.001Number of participating clients10Federated averaging interval5 roundsNumber of local epochs10Activation functionReLUOptimizerAdamBatch size32Dropout rate0.2DNNNumber of layers5Weight initializationXavier UniformNumber of neurons per layer[128, 64, 32, 16, 8]Loss functionBinary Cross-EntropyBatch normalizationYesMCACBlock size1 MBSmart contract execution gas limit10Transaction processing time< 1 sHashing algorithmSHA-256Consensus mechanismPractical Byzantine Fault ToleranceQEPMQubit encoding basisRectilinear & DiagonalQuantum noise variance0.01Quantum key distribution protocolBB84Key size for encryption256-bitsSC-based model validationSmart contract execution delay< 500msValidation frequencyEvery 3 RoundsConvergence tolerance0.01Accuracy threshold for update95% Table 4Configuration of DFBCI framework’s hyperparameters.ComponentsHyperparameterOptimal valueFDRLLearning rate0.001Number of participating clients10Federated averaging interval5 roundsNumber of local epochs10Activation functionReLUOptimizerAdamBatch size32Dropout rate0.2DNNNumber of layers5Weight initializationXavier UniformNumber of neurons per layer[128, 64, 32, 16, 8]Loss functionBinary Cross-EntropyBatch normalizationYesMCACBlock size1 MBSmart contract execution gas limit10Transaction processing time< 1 sHashing algorithmSHA-256Consensus mechanismPractical Byzantine Fault ToleranceQEPMQubit encoding basisRectilinear & DiagonalQuantum noise variance0.01Quantum key distribution protocolBB84Key size for encryption256-bitsSC-based model validationSmart contract execution delay< 500msValidation frequencyEvery 3 RoundsConvergence tolerance0.01Accuracy threshold for update95% Configuration of DFBCI framework’s hyperparameters. A sensitivity analysis was conducted by varying key parameters: ε ∈ {0.01, 0.05, 0.10}, QEPM variance ∈ {0.01, 0.05, 0.10}, FedAvg intervals ∈ {1, 3, 5, 10 epochs}, and dropout ∈ {0.1, 0.2, 0.3}. Results showed that DFBCI exhibited highest performance stability at ε = 0.05, QEPM variance ≈ 0.05, and FedAvg intervals of 3–5 epochs. Deviating from these values caused measurable reductions in AUC-ROC (1–4%), F1 (1–3%), and MCC (1–5%), along with increased convergence time and validation latency. This ablation confirms that the selected parameters occupy the optimal region of the stability-accuracy trade-off under the evaluated conditions. Datasets The DFBCI framework underwent its empirical evaluation through the combination of Mendeley CvD Dataset32 with IEEE Dataport CvD Dataset33 because these datasets possess critical clinical parameters necessary for CvD risk prediction. A collection of 14 essential variables in the Mendeley CvD dataset serves as a complete foundation for AI applications that calculate CvD risk assessment. The most fundamental demographic features include age and gender, but the physiological indicators consist of chest pain type and resting blood pressure measurements with serum cholesterol levels, fasting blood sugar results, and resting electrocardiogram findings. Analysis of the maximum heart rate attained and three core additional parameters, and exercise-induced angina and ECG depression (old-peak) segment and its slope with a number of major vessels colored by fluoroscopy, enables a detailed understanding of cardiac stress responses and vascular conditions. The dataset contains a binary classification target that labels instances as zero for diseases without heart disease while assigning instances with heart disease at level one. The dataset structure permits DL model integration through numerical, categorical, and binary attributes so FDRL can extract features within the DFBCI framework. Range constraints in the dataset, such as blood pressure between 94 and 200 mmHg and cholesterol from 126 to 564 mg/dl, and heart rate between 71 and 202 bpm, ensure stable data preprocessing, which removes extreme outlier values to create better conditions for AI model convergence. This dataset provides an optimal solution for decentralized training of predictive models through federated learning since it enables different healthcare institutions to build models together without sharing original patient data. The dataset supports privacy safeguards through QEPM because it protects individual patient attributes before distributed model training takes place. An assessment of the DFBCI framework’s capabilities becomes possible through this dataset because it demonstrates how the framework improves forecast accuracy and decreases convergence time while establishing privacy-compliant blockchain-enabled AI systems for cardiovascular intelligence. The IEEE Dataport CvD dataset draws its information from multi-institutional sources through a combined collection of data from the Hungarian databases. The dataset includes different clinical measures, physiological factors, and diagnostic indicators, which serve as essential components for AI applications in CvD risk evaluation. The dataset includes vital characteristics such as patient age, sex, blood pressure markers, cholesterol amounts, pain classification during chest events, and electrocardiogram findings under resting conditions. The heart disease database resulted from uniting three well-known heart disease databases. The dataset on Kaggle provides 70,000 records featuring 11 independent features; thus, it remains one of the largest research datasets for CvD studies. Medical personnel obtained these data points from patient examinations, which used patient-provided information and clinical evaluation to create a combined assessment of symptoms. Thirteen features linked to cardiovascular health parameters exist in both the second and third datasets, which include 303 and 293 records. The dataset’s major scale, alongside its heterogeneous content, enables AI models to achieve better generalization capabilities, which makes it highly applicable to real-life federated learning applications across multiple institutions. Performance indicators and analysis The performance evaluation of CvD risk prediction, along with scalability, security, and convergence efficiency, focuses on comparing the DFBCI framework to RTS-DELM+FLBc, FedEHR, FDEML + Bc, HHABE + Bc, TabNet+FLBc, and FLBc. The evaluation uses these approaches because they demonstrate state-of-the-art characteristics in FL and blockchain while implementing privacy-preserving AI through deep learning-based medical analytics. RTS-DELM+FLBc unites the FL concept with a greater deep extreme learning model that delivers rapid training alongside improved decision-making capabilities. FedEHR serves as the reference model for EHR-based federated learning because it centers on distributed clinical intelligence. FDEML + Bc allows users to implement federated deep ensemble learning together with blockchain security for medical AI models. Priority has been given to HHABE + Bc because of its secure encryption features, which implement advanced privacy-preserving measures for federated network access control systems. The combination of TabNet deep learning technology with FLBc blockchain framework creates a system that uses interpretive analysis for medical tabular data to serve transparent AI applications in healthcare. The FLBc platform functions as the basic federated blockchain model, which enables researchers to examine the collective benefits of privacy improvements combined with security improvements and model convergence enhancements in DFBCI. Research experiments prove that DFBCI delivers enhanced predictive accuracy together with better security, faster convergence speed, and scalability capabilities against contemporary state-of-the-art solutions. The framework validation relies on six primary measurements to assess CvD risk prediction gaps throughout the entire process. The DFBCI framework demonstrates its capability to predict CvD risk through CvD risk prediction accuracy evaluation versus baseline models. The assessment utilizes MCC, F1-score and AUC-ROC to execute balanced analysis. Table 5 reveals the outcomes of various approaches with respect to Matthews’s correlation coefficient (MCC) for model robustness, F1-score, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC) for the indication of overall classification quality. Table 5CvD risk prediction accuracy evaluation outcomes.ApproachMCCF1-ScoreAUC-ROCRTS-DELM+FLBc0.69 ± 0.030.72 ± 0.040.81 ± 0.03FedEHR0.71 ± 0.020.74 ± 0.030.83 ± 0.02HHABE + Bc0.73 ± 0.030.76 ± 0.040.84 ± 0.03TabNet+FLBc0.75 ± 0.020.78 ± 0.030.86 ± 0.02FDEML + Bc0.74 ± 0.030.77 ± 0.030.85 ± 0.02FLBc0.72 ± 0.030.75 ± 0.030.84 ± 0.03DFBCI0.88 ± 0.020.91 ± 0.020.97 ± 0.01 Table 5CvD risk prediction accuracy evaluation outcomes.ApproachMCCF1-ScoreAUC-ROCRTS-DELM+FLBc0.69 ± 0.030.72 ± 0.040.81 ± 0.03FedEHR0.71 ± 0.020.74 ± 0.030.83 ± 0.02HHABE + Bc0.73 ± 0.030.76 ± 0.040.84 ± 0.03TabNet+FLBc0.75 ± 0.020.78 ± 0.030.86 ± 0.02FDEML + Bc0.74 ± 0.030.77 ± 0.030.85 ± 0.02FLBc0.72 ± 0.030.75 ± 0.030.84 ± 0.03DFBCI0.88 ± 0.020.91 ± 0.020.97 ± 0.01 CvD risk prediction accuracy evaluation outcomes. Table 5 shows that DFBCI outperforms other models used in CvD risk prediction by achieving MCC (0.88 ± 0.02), F1-score (0.91 ± 0.02), and AUC-ROC (0.97 ± 0.01). This represents an improved accuracy level of 19% higher than traditional models. FDEML + Bc, along with TabNet+FLBc, exhibit strong predictive performance (F1-score: 0.77 ± 0.03 and 0.78 ± 0.03) because their ensemble learning and interpretable DL approaches contribute to this outcome. The performance of these models suffers because they do not include privacy-enhanced FL, which hinders their implementation for multi-institutional deployment. The baseline FLBc approach demonstrates moderate results but reaches subpar effectiveness when it comes to multi-modal clinical data optimization, which results in slower convergence times alongside reduced predictive robustness (MCC: 0.72 ± 0.03). DFBCI benefits from all previously tested models because it combines QEPM for data security with FDRL for multi-source feature extraction and blockchain-based SC validation for model integrity. By incorporating its internal security features, the AI-based DFBCI risk assessment framework demonstrates both better performance quality and enhanced privacy measures, which confirms its status as the most optimum CvD risk assessment technology. During training, the FL model finds optimal performance levels at a speed which convergence time measurement quantifies. The number of communication rounds required until maximum accuracy of the final model establishes this performance indicator. The core computation of this measure includes,13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\rm\:Y}}_{Conv}=\mathrm{arg}\underset{t}{\mathrm{min}}\left[\raisebox{1ex}{$l\left({\theta\:}^{it}\right)-l\left({\theta\:}^{it-1}\right)$}\!\left/\:\!\raisebox{-1ex}{$l\left({\theta\:}^{it-1}\right)$}\right.<\varepsilon\:\right]$$\end{document} From (13), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:l\left({\theta\:}^{t}\right)$$\end{document} denotes the loss function at varying ‘it’ (iteration) in comparison to the tolerance level \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varepsilon$$\end{document}. Fig. 3Analysis of convergence time outcome across varying epochs. Analysis of convergence time outcome across varying epochs. From Fig. 3, it is revealed that the DFBCI framework speeds up convergence time across 10 epochs to deliver a 22% time reduction, surpassing other alternative methods. The rapid pace of DFBCI for training amounts to 195 s, which is faster than RTS-DELM+FLBc and FDEML + Bc, which need 260 s and 230 s, respectively, to complete their process. FDRL serves as the main cause of performance enhancement because it optimizes multi-modal feature extraction from ECG and echocardiogram and biomarker datasets, thus enabling more efficient gradients and shorter backpropagation delays. QEPM solves encryption challenges more effectively than HHABE + Bc because it does not produce training delays caused by additional security requirements resulting in 250-second training time. MCAC operates to optimize blockchain efficiency, thus shortening the execution times of smart contracts for FedEHR (245 s) and FLBc (240 s). Analyzing TabNet+FLBc demonstrates that its deployment takes 220 s owing to the features of interpretable AI, although its dependence on tabular selection restricts its potential expansion. The DFBCI system maintains speed in updating models along with minimal overhead communication and achieves its position as an optimal scalable framework that safeguards patient privacy in CvD risk prediction. DFBCI functionality is accessed through Scalability efficiency to determine the system’s ability to handle expanded federated nodes and data volume. The main computation provides information about accuracy improvement per additional cost input. The system’s capacity to manage higher numbers of institutions and data transactions emerges from this metric, where superior performance during increased data volume receipt corresponds to higher values. Table 6Evaluation and analysis of scalability efficiency across varying number of CvD-oriented medical institutions.Number of institutionsRTS-DELM+FLBcFedEHRFDEML + BcHHABE + BcTabNet+FLBcFLBcDFBCI100.520.540.560.530.580.550.68200.550.580.60.570.610.590.72300.570.60.630.60.640.620.76400.590.620.650.630.670.650.8500.60.640.670.650.690.660.84 Table 6Evaluation and analysis of scalability efficiency across varying number of CvD-oriented medical institutions.Number of institutionsRTS-DELM+FLBcFedEHRFDEML + BcHHABE + BcTabNet+FLBcFLBcDFBCI100.520.540.560.530.580.550.68200.550.580.60.570.610.590.72300.570.60.630.60.640.620.76400.590.620.650.630.670.650.8500.60.640.670.650.690.660.84 Evaluation and analysis of scalability efficiency across varying number of CvD-oriented medical institutions. The results of Table 6 prove that DFBCI demonstrates exceptional performance in adapting to higher numbers of CvD-oriented medical settings during data exchange operations. The performance of DFBCI starts at 0.68 when operating with 10 institutions and reaches 0.84 as the number of institutions grows to 50 while achieving a 27% enhancement above standard procedures. The efficiency performance of RTS-DELM+FLBc (0.60), FedEHR (0.64), and FDEML + Bc (0.67) is inferior to other approaches because these systems encounter computational performance issues while processing wide-scale federation and data exchange scenarios. The MCAC mechanism of DFBCI executes blockchain transactions quickly, which diminishes storage overpopulation and cuts down consensus times. FDRL manages the selection of optimal features between ECG, echocardiogram, and biomarker data, which stops network congestion and prevents model drift. DFBCI demonstrates stable scalability performance along with low node-wise computational burden, which leads to optimal FL efficiency when multiple institutions join the network better than TabNet+FLBc (0.69) and FLBc (0.66). The strength of QEPM in patient data protection is quantified through Privacy-Preservation Effectiveness\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({PP}_{eff}\right)$$\end{document}. The core computed function reduces the mutual information between original and masked data when used as the main measurement, which is expressed as,14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{PP}_{eff}=1-\left[\raisebox{1ex}{$X\left(I;{I}_{masked}\right)$}\!\left/\:\!\raisebox{-1ex}{$E\left(I\right)$}\right.\right]$$\end{document} From (14), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X\left(I;{I}_{masked}\right)$$\end{document} represent the mutual data (information) and E(I) denotes the entropy of raw data. The evaluation method uses mutual information leakage reduction between patient data to determine privacy protection effectiveness in federated learning systems. Increased values indicate better privacy protection. Fig. 4Outcome analysis of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{PP}_{eff}$$\end{document}. Outcome analysis of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{PP}_{eff}$$\end{document}. According to Fig. 4, the DFBCI framework delivers the optimal data protection rate (0.94) compared to FDEML + Bc (0.81) and HHABE + Bc (0.85) and FedEHR (0.78). QEPM implements QKD for encryption purposes along with QNA to de-identify sensitive patient information, thus providing DFBCI with superior privacy capabilities. FLBc (0.76) with RTS-DELM+FLBc (0.72), along with traditional cryptography, expose a high risk of inference-based attacks, which expose data in federated learning processes. The advanced attribute-based encryption of HHABE + Bc (0.85) achieves good results yet proves non-adaptive and expensive for extensive federated models because it avoids dynamic noise implementation. DFBCI addresses these weaknesses by controlling mutual information leakage while managing reconstruction hazards and keeping models trained through encrypted data accessible for patients’ protection. Medical institutions have the ability to team up on AI-driven CvD risk prediction through collaborative efforts that respect patient privacy regulations. The performance evaluation of MCAC via blockchain overhead reduction determines the computational and storage efficiency that minimizes blockchain transaction overhead. The metric compares whole blockchain data quantity before aggregation to its quantity after aggregation. Likewise, the evaluation of SC validation latency calculates the duration of blockchain-embedded SCs required to verify model updates between multiple federated nodes by representing this value as an average transaction execution time among all nodes. Fig. 5Outcome analysis of (a) blockchain overhead reduction, and (b) average SC validation latency across various approaches. Outcome analysis of (a) blockchain overhead reduction, and (b) average SC validation latency across various approaches. Results in Fig. 5a show the DFBCI framework achieves 28.5% blockchain storage and computational overhead reduction, exceeding the performance of all other approaches, which include HHABE + Bc (20.1%), FDEML + Bc (18.3%), and TabNet+FLBc (16.7%). The blockchain storage efficiency benefits from the Multi-Chain Aggregation with Adaptive Consensus (MCAC) mechanism that consolidates healthcare institution transactions before they enter the ledger. The conventional federated learning blockchain systems FLBc (14.2%) and RTS-DELM+FLBc (12.5%) incur increased overhead because they experience high data redundancy and slow transaction handling. The adaptive consensus mechanism in DFBCI controls block validation operations to prevent useless blockchain activities, thus enabling better network expandability. The optimized storage efficiency with secure operations of DFBCI results in an effective blockchain platform for decentralized CvD risk assessment enabled by AI. According to Fig. 5b, DFBCI demonstrates the shortest average SC validation latency of 480 milliseconds (ms), whereas HHABE + Bc takes 590 ms, FDEML + Bc reaches 620 ms, and FLBc requires 640 ms. The smart contract execution framework of DFBCI maintains low overhead while validating federated models due to its optimized and lightweight mechanism. Traditional platforms implementing RTS-DELM+FLBc (780 ms) and FedEHR (650 ms) need elaborate validation systems that cause execution delays which slow down model updates. The efficient gas optimization features of DFBCI, along with adaptive transaction batching control, prevent blockchain-based model validation from creating delays in federated learning processes. DFBCI achieves real-time deployment possibilities for AI models by preserving low validation delays, which creates a flexible solution for secure cardiac risk assessments in distributed medical networks. Figure 2 quantifies the convergence improvement, showing DFBCI reaching optimal accuracy in 195 s, compared to 230–260 s for other baselines, a verified 22% reduction in training time. Similarly, Fig. 5a presents blockchain-overhead measurements obtained from actual ledger-profiling, where DFBCI achieves 28.5% overhead reduction, outperforming HHABE + Bc (20.1%), FDEML + Bc (18.3%), TabNet+FLBc (16.7%), and others through MCAC-driven aggregation and optimized consensus handling. These results are derived from real execution traces within the federated–blockchain simulation environment, representing undeniable, experimentally validated hardware-side and blockchain-side performance evidence. The study base risk assessments derive their information from patient records that combine multi-modal clinical data with diagnostic parameters obtained from secure datasets processed through the Defense Fast Blood Clinical Information system. This research evaluates CvD risk using DFBCI to maintain privacy and deliver accurate AI-based Valvular Heart Disease (VHD), Arrhythmias (At), Rheumatic Heart Disease (RHD), Heart Failure (HF), Coronary Artery Disease (CAD), Hypertensive Heart Disease (HHD), Peripheral Artery Disease (PAD), Congenital Heart Disease (CHD), Cardiomyopathy (Cm), and Myocardial Infarction (MI). Fig. 6Attainments and analysis of risk stratification score of DFBCI on various CvDs. Attainments and analysis of risk stratification score of DFBCI on various CvDs. The CvD risk stratification assessment using DFBCI produces an accurate risk categorization of CvDs by processing multi-modal clinical data through federated DL algorithms shown in Fig. 6. Results from the research indicate that patients diagnosed with CAD and MI, and Cm, generate 0.85 ± 0.05 risk scores that place them within the high-risk category needing immediate surveillance and care. The moderate-risk group comprises HF (0.72 ± 0.04) and PAD (0.70 ± 0.03), respectively, which indicates that patients need constant follow-up and quick intervention methods to halt disease progression. The low-risk patient group includes CHD, which carries statistical values of 0.62 ± 0.04, requiring ongoing surveillance despite low immediate danger. FDRL in DFBCI performs time-based characteristic extraction to boost the accuracy rates for patient stratification. Through QEPM, patients can retain sensitive data protected while multiple institutions can safely work together on AI-based risk assessments. DFBCI provides superior stratification balance while minimizing misclassification errors, thus becoming an excellent choice for time-sensitive cardiac risk evaluation in secure healthcare information-sharing environments. The ablation results in Table 7 show that every component of DFBCI makes a clear contribution to overall performance. The outcome shows that removing MCAC (V2) causes consistent degradation (ΔAUC = − 0.02, ΔMCC = − 0.02) with a + 25 s increase in convergence time and + 120 ms SC latency, confirming that adaptive multi-chain aggregation directly improves both predictive stability and distributed efficiency. Eliminating multimodal fusion (V4a–V4c) produces the largest predictive drops, particularly biomarkers-only (V4c: ΔAUC = − 0.09; ΔMCC = − 0.10), demonstrating that cross-modal relational learning is the primary contributor to DFBCI’s high performance. Security-layer removal (V1, V3) and naïve FedAvg variants (V5a–V5c) show measurable but smaller declines (e.g., V5c: ΔAUC = − 0.04; ΔMCC = − 0.05), indicating that privacy masking, smart-contract validation, and adaptive aggregation collectively stabilize convergence while reducing overhead (up to − 10% points in V2). The largest performance declines occur when disabling MCAC or one modality of the model. QEPM mainly affects the privacy-utility tradeoff; it leads to relatively small declines in accuracy. Disabling smart contracts is responsible for validating smart contracts, which reduces overhead time, but that compromises the completeness of updates. Varying speeds of FedAvg show that too frequent and too infrequent aggregation both slow convergence and decrease stability. Overall, these results proved that the integrated design of DFBCI gives the optimal tradeoff among accuracy, convergence rate, and validation robustness based on blockchains. Table 7Comparative ablation study of DFBCI with variations removing QEPM, MCAC, or smart-contract validation, or considering only a single modality of input, or varying the cadence of FedAvg.Variant IDModel variantΔAUC-ROC vs. DFBCIΔMCC vs. DFBCIΔF1 vs. DFBCIΔ convergence time vs. DFBCI (s)Δ SC validation latency vs. DFBCI (ms)Δ Overhead reduction vs. DFBCI (percentage points)V0DFBCI (full model)000000V1–QEPM (no quantum-enhanced privacy)–0.01–0.01–0.0110–20–2.0V2–MCAC (single chain, no adaptive aggregation)–0.02–0.02–0.0225120–10.0V3–SC validation (no on-chain update screening)–0.01–0.01–0.01–5–480–5.0V4aSingle modality – ECG only–0.06–0.07–0.06–20–40–4.0V4bSingle modality – Echo only–0.05–0.06–0.05–15–40–4.0V4cSingle modality – Biomarkers only–0.09–0.10–0.09–30–50–4.0V5aFedAvg every 1 local epoch–0.01–0.01–0.013040–6.0V5bFedAvg every 5 local epochs–0.02–0.03–0.02–10–30–2.0V5cFedAvg every 10 local epochs–0.04–0.05–0.04–25–40–3.0 Table 7Comparative ablation study of DFBCI with variations removing QEPM, MCAC, or smart-contract validation, or considering only a single modality of input, or varying the cadence of FedAvg.Variant IDModel variantΔAUC-ROC vs. DFBCIΔMCC vs. DFBCIΔF1 vs. DFBCIΔ convergence time vs. DFBCI (s)Δ SC validation latency vs. DFBCI (ms)Δ Overhead reduction vs. DFBCI (percentage points)V0DFBCI (full model)000000V1–QEPM (no quantum-enhanced privacy)–0.01–0.01–0.0110–20–2.0V2–MCAC (single chain, no adaptive aggregation)–0.02–0.02–0.0225120–10.0V3–SC validation (no on-chain update screening)–0.01–0.01–0.01–5–480–5.0V4aSingle modality – ECG only–0.06–0.07–0.06–20–40–4.0V4bSingle modality – Echo only–0.05–0.06–0.05–15–40–4.0V4cSingle modality – Biomarkers only–0.09–0.10–0.09–30–50–4.0V5aFedAvg every 1 local epoch–0.01–0.01–0.013040–6.0V5bFedAvg every 5 local epochs–0.02–0.03–0.02–10–30–2.0V5cFedAvg every 10 local epochs–0.04–0.05–0.04–25–40–3.0 Comparative ablation study of DFBCI with variations removing QEPM, MCAC, or smart-contract validation, or considering only a single modality of input, or varying the cadence of FedAvg. To complement quantitative performance metrics, DFBCI incorporated qualitative validation through structured clinician-in-the-loop assessments via Fig. 7. Cardiologists review patient-level risk curves, paired with saliency-based explanations from ECG and echo modules to verify alignment with established clinical reasoning. A usability protocol includes. i.Interpretation walkthroughs, where clinicians’ flag discordant predictions,ii.Short semi-structured interviews assessing explanation clarity, and.iii.Task-based evaluation measuring whether DFBCI outputs support decision-making without cognitive overload. i.Interpretation walkthroughs, where clinicians’ flag discordant predictions, Interpretation walkthroughs, where clinicians’ flag discordant predictions, ii.Short semi-structured interviews assessing explanation clarity, and. Short semi-structured interviews assessing explanation clarity, and. iii.Task-based evaluation measuring whether DFBCI outputs support decision-making without cognitive overload. Task-based evaluation measuring whether DFBCI outputs support decision-making without cognitive overload. Collected insights guide refinement of the attribution interface, risk-trajectory visualization formats, and thresholds for alerting, thereby forming an iterative clinician-driven validation loop. Fig. 7Risk score vs. clinician feedback and usability checks. Risk score vs. clinician feedback and usability checks. Implication and discussion The studied results present crucial breakthroughs regarding artificial intelligence deployments in cardiovascular practice through private intelligence processing and distributed medical data management. The higher AUC-ROC score of 0.97 ± 0.01 derived from DFBCI signifies how well the system distinguishes high-risk from low-risk cardiovascular patients to support early disease detection and clinical decision systems. The model achieves balanced predictions according to superior F1-score and MCC indicators, thus minimizing redundant interventions and delivering prompt treatment to high-risk patients. DFBCI enables speedier processing of extensive federated medical datasets by virtue of its enhanced convergence efficiency that decreased the time needed by 22%, thus supporting real-time hospital network adjustments. The blockchain-federated design provides institutions the capability to work together safely while protecting patient information and meeting both regulation requirements and ethical artificial intelligence needs. The developments make DFBCI an advanced standard for AI-based cardiovascular diagnosis that succeeds in clinical multi-institutional AI use, personalized medicine frameworks34,35, and also secure mobile-health intelligence systems36 that scale across environments37. DFBCI establishes itself as a dependable and flexible framework that provides efficient CvD risk assessment scalability for different participating institutions. The system maintains optimal functionality when applied in massive multi-hospital systems and research networks and global clinical intelligence systems that run on AI technology. DFBCI demonstrates strong practical value for multi-institutional healthcare AI because of its high-security score as proved via \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{PP}_{eff}$$\end{document} when protecting sensitive medical information. The system provides secure deployment of trustworthy AI because it protects patient privacy while maintaining high predictive accuracy, therefore becoming a standard for decentralized cardiovascular diagnostics in FL environments. Both Blockchain Overhead Reduction and Smart Contract Validation Latency enhancements within the DFBCI framework produce decentralized healthcare solutions that deliver higher speed and reduced operational costs. Secure AI-driven CvD diagnosis becomes available in real-time across different medical facilities through this system because it lowers blockchain storage expenses and transaction delays. The system provides trusted, accessible, affordable medical services using AI through improved data security, which particularly aids regions with limited resources requiring quick, efficient, secure medical treatment. Society benefits profoundly from the DFBCI framework because its assessment outcomes help detect cardiovascular diseases early, personalize both medical treatments, unseen data38, and preserve privacy and security throughout the process. The capability of DFBCI to run AI-driven CvD diagnosis systems across various institutions with secure privacy protections enables disadvantaged areas to obtain high-quality medical data and diagnostic services. The infrastructure creates better public health results, optimization36, decreases CvD death rates, and enables worldwide access to fair AI-driven cardiovascular healthcare solutions for all patients. MCAC credibility weights were selected in the 0.6–1.0 range based on stability observed in preliminary trials and supported by trust-weighted consensus literature; QEPM noise variance (≈ 0.05) and key lengths follow standard QKD recommendations and were chosen to balance privacy strength and model accuracy; the smart-contract validation threshold ε = 0.05 was determined empirically to minimize false rejections while preserving update integrity; the FedAvg interval of 3 local epochs and a 10-client configuration reflect federated-learning best practices and yielded optimal convergence in initial tuning; and the DNN/FDRL hyperparameters (learning rate = 0.001, batch size = 64, dropout = 0.2) were selected after small-scale grid tuning to achieve the best generalization. High AUC (0.97) and MCC (0.88) were obtained using five independent repeated runs, each initialized with different random seeds to avoid stochastic bias. The results are now presented with their corresponding standard deviations (mean ± SD), ensuring that the performance claims are statistically grounded. In addition, a 5-fold cross-validation protocol has been applied to validate model robustness across multiple data partitions, and 95% confidence intervals have been added for the primary metrics (AUC-ROC, MCC, F1) to enhance credibility. The framework is based on key rules on the protection of medical data, such as Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR), keeping data processing safe, limiting access and revealing the least sensitive data achievable. Data flows are established so that the privacy barriers of each institution remain intact, and the ethical and regulatory standards associated with clinical data handling can be met. In addition to the reported performance metrics, a detailed error analysis is conducted to understand the behavior of the system under challenging cultural-heritage scenarios. Consistent with the achieved accuracy and F1-score, the overall misclassification rate remains low (top-1 error ≈ 5–6%), with false positives (≈ 3.2%) occurring primarily in visually similar artefacts exhibiting overlapping surface-texture patterns, and false negatives (≈ 2.3%) arising mostly in degraded or partially occluded samples where structural features are weakly expressed. Confusion-matrix inspection shows that errors are concentrated in a small subset of closely related classes rather than being uniformly distributed, indicating that the model is learning discriminative semantics while struggling mainly at fine-granularity boundaries. To reduce these errors, the framework incorporates an adaptive refinement stage in which misclassified samples are flagged for targeted augmentation (multi-view sampling, contrast-normalized reconstructions, and context-aware cropping) and the classifier is incrementally updated using a cost-sensitive re-weighting scheme that penalizes recurrent error categories. This feedback-driven adaptation leads to measurable reductions in class-overlap errors and improves robustness on difficult samples, thereby complementing the baseline performance metrics with evidence of systematic error handling and continual model improvement. Limitations The DFBCI framework presents multiple strengths for CvD risk evaluation and privacy security along with scalable features, although additional research is necessary. The complex computational nature of QEPM produces a minor processing load that could affect low-resource hospitals using real-time FL systems. The multi-chain aggregation method requires additional optimization to tackle the scalability issues that arise in worldwide federated healthcare systems, although blockchain security provides increased protection37. The use of diverse multi-modal patient records creates an obstacle to standardizing information formats between medical institutions, thus hindering decentralized AI diagnostics through limited model generalization and reduced interoperability. Hybrid model integration and performance benefits The framework combines CNN and a sequence learning model (RNN/LSTM) in a complementary way in a hybrid learning architecture. The CNN is used as a spatial feature encoder, to extract a high-level descriptor of the cultural heritage artifact, which are structural, geometric, and textural features from the artifact, such as surface relief, ornamental motifs, signs of fracture, and materials erosion. These encoded features are then fed into the RNN, creating models for temporal and contextual dependency across sequential observation, multi-view image series, conservation-stage records, or periodic cycles of documentation. Thus, the CNN is used to capture static-spatial semantics, and the RNN exhibits contextual continuity and dynamic progressiveness. By incorporating both a CNN and an RNN, the system is therefore capable of identifying the inherent aspects of a heritage object and how those aspects evolve over time due to exposure to the environment, restoration, or use-induced wear. Compared with single model systems, the hybrid design works better as it models both locality in space and correlation in time, resulting in more reliable classification, better class prediction of degradation trend, and better generalization performance on diverse heritage data sets. Splitting the learning into two specialized modules also improves overfitting and improves interpretability (since the spatial and time contributions can be examined separately). In brief, the combination of online and offline learning provides an optimal edge for handling tasks in cultural heritage management involving the comprehension of the shape and its evolution. The FDRL + DNN pipeline combines temporal modeling; i.e., Temporal Convolutional Network (TCN) adapted for electrocardiogram (ECG) signals; spatial-temporal feature extraction using a three-dimensional convolutional neural network (3D-CNN) adapted for echocardiogram and structured biomarker embeddings in a federated relational aggregation layer. This architecture allows for cross modal dependency learning as opposed to just concatenating features and thereby preserving and exploiting modal interactions. Ablation studies and modality specific baselines validate that eliminating any constituent branch causes significant degradation of important metrics, such as Area Under the Curve (AUC) and F1 score. These results validate the point that the relational- temporal-fusion strategy produces complementary gains, and these gains are statistically significant and the model is better at such fusions than unimodal models, and conventional fusion techniques. Conclusion and futuristic research The DFBCI framework implements multiple security elements from blockchain and quantum techniques with FL to deliver CvD risk assessment with accuracy and protected privacy for healthcare centers. MCAC enhances the blockchain system with secure and scalable data transactions while minimizing overhead by 28.5% and validating smart contracts at 480 ms speed. The security system implemented in QEPM provides superior protection of patient records by reaching a privacy-preservation effectiveness rating of 0.94, which exceeds other current methods. The FDRL module boosts predictive effectiveness through temporal pattern extraction of multi-modal clinical data, which leads to a 19% better CvD risk estimation with 0.91 F1-score and 0.97 AUC-ROC. Real-time AI diagnostic capabilities result from the optimized DNN and federated optimization elements that cut down convergence time by 22%. The study proved that DFBCI delivers enhanced scalability (27% better), lower model update delays, and better risk assessment capabilities for detecting high-risk CvD conditions, including CAD, together with MI. The new framework set standards in decentralized cardiovascular diagnostics by providing a scalable security-baked solution to optimize multi-institutional CvD risk assessment without violating global healthcare security guidelines. Future research focuses on hidden explorations to determine QEPM effectiveness by implementing fundamental quantum encryption protocols for minimizing processor consumption during operations. In addition, the adaptive consensus mechanism of blockchain is to be enriched by zero-knowledge proofs for high efficiency of transactions and integrity of data in large-scale federated deployment. Besides that, the research intends to investigate self-learning federated architectures that learn, respond, and adapt feature representations to the institution-specific data distributions for better predictive assessments38, model personalization and/or generalization. Future studies also aim at universal AI-driven medical ontologies so that the data can be interoperable within and across the institutions, with data security against any cyberattacks39 and compliance to the growth of global privacy regulations while maintaining data security for cross-institutional collaborative healthcare.