Design and development of an mHealth application for pressure ulcer care and caregiver support Authors: Shreenidhi Jogi, Vishal Shanbhag, Lakshay Chauhan, Siddhartha Chhauda, Utkarsh Dubey, Ajitha K. B. Shenoy, Elsa Sanatombi Devi Published: 2025 ABSTRACT IntroductionSmartphone accessibility has enabled the widespread use of mobile health applications for managing health conditions. While mobile technology is increasingly adopted globally, integrated digital solutions specifically supporting home-based pressure ulcer care remain limited. This study aimed to design and develop a mobile health (mHealth) application named IPI (Interprofessional Pressure Injury) application that integrates artificial intelligence-based pressure ulcer staging, caregiver-focused education, personalized nutritional support, and visual wound monitoring to assist caregivers and healthcare professionals in delivering timely and effective care.MethodsA comprehensive deep learning framework was developed using a clinically validated dataset of pressure ulcer images spanning six categories, including healthy tissue and Stage 1–4 ulcers. To address class imbalance and subtle inter-class variability, a class-adaptive augmentation pipeline and an enhanced Vision Transformer architecture with hierarchical feature representation and specialized self-attention were implemented. Training employed a stratified 5-fold cross-validation, class-balanced focal loss, regularization techniques, and a two-tiered ensemble strategy.ResultsThe proposed k-fold ensemble model achieved an accuracy of 0.9705 and macro F1 score of 0.9695, with perfect classification of Stage 4 ulcers and substantial improvements for underrepresented classes.DiscussionThese results demonstrate the model's effectiveness for pressure ulcer classification, offering a robust foundation for real-time clinical decision support. The application supports remote monitoring, healing status detection, and educational access, especially in resource-limited settings. This holistic solution not only enhances caregiver confidence and independence but also aids clinicians in wound assessment and intervention planning. A future experimental study will validate the app's clinical utility, impact on patient outcomes, and potential to improve the quality of home-based pressure ulcer management. Introduction Pressure ulcers are defined as “localized damage to the skin and underlying soft tissue usually over a bony prominence or related to a medical or other device” (1). Pressure ulcers result from a combination of intrinsic factors, such as reduced mobility, aging skin, comorbidities, and poor nutrition, along with extrinsic factors including pressure, shear, friction, and moisture. They may present as either intact skin or open wounds, often accompanied by pain, and typically arise due to intense and/or prolonged pressure, or pressure combined with shear. The tolerance of soft tissue to these forces is further influenced by factors such as microclimate, nutritional status, tissue perfusion, and overall health (1, 2). The National Pressure Injury Advisory Panel (NPIAP) classifies pressure injuries into distinct stages: stage 1, stage 2, stage 3, stage 4, unstageable pressure injuries, deep tissue pressure injuries, and mucosal membrane pressure injuries (1). The incidence of pressure ulcers has been reported to range from 1.9% to 35%, while prevalence estimates vary between 11.1% and 31% across different care settings and populations (3). In 2019, an estimated 3.17 million new cases of pressure ulcers were reported globally, with the majority occurring in individuals aged over 75 years (4). The burden of pressure ulcers has been steadily increasing due to the aging global population. Therefore, early diagnosis of pressure ulcers and effective treatment based on accurate diagnostic results are crucial to preventing disease progression and reducing the long-term burden of pressure ulcers (5). The details of a pressure ulcer must be carefully documented when it develops. Important characteristics to document include its stage, size, location, the presence of eschar, type and amount of exudate, any odour, and indications of infection. This detailed documentation is vital for effective treatment planning and progress evaluation of pressure injuries (2, 6). The initial assessment of pressure ulcers is typically conducted by physicians or nurses in hospitals or clinics, with ongoing follow-up care provided daily or weekly by community nurses or other healthcare providers. However, in-person evaluation by specialists is not always feasible, particularly for patients lacking access to specialized transportation, those without family support, or those living in remote areas. Due to which, many patients with pressure ulcers are managed by informal home caregivers, which can increase the risk of wound infection, delayed healing, and hospital admissions (7, 8). Despite significant progress in wound care, the assessment of pressure ulcers remains predominantly dependent on visual inspection and the clinical judgment of healthcare providers. This reliance introduces substantial inter-assessor variability, leading to inconsistencies in diagnosis and staging (9). Such diagnostic subjectivity can result in inappropriate treatment selection, delays in timely intervention, and diminished patient outcomes, with broader implications for care quality and healthcare expenditure. Telemedicine, telehealth, and eHealth refer to the remote application of information and communication technologies to facilitate healthcare delivery. These technologies assist in the diagnosis, treatment, monitoring, and prevention of diseases or injuries. These technologies also help in health education for caregivers which enhances the well-being of individuals and communities (10, 11). Smartphones have become more accessible for the people. The overall usage of smartphones for various purposes has increased worldwide. This also has contributed for the use of mobile applications for managing health conditions, disease and injuries. Mobile health (mHealth) refers to the utilization of mobile applications (apps), wearable devices, and mobile technology to deliver medical information, collect or access health data, offer clinical and personal services, and support healthcare delivery across both clinical and non-clinical environments (12). The role of mHealth apps in healthcare is significantly growing. The mHealth apps enable physicians and healthcare personnel to remotely assess, monitor and analyse the progress of the pressure ulcers. This indeed improves pressure ulcer evaluation in clinical, hospital and home-based settings. The use of mHealth apps for various health conditions is gradually increasing. It is easy to update the information and resources in smartphones with the help of an internet connection as compared to the traditional paper-based methods. Additionally, mHealth apps enable users to develop skills, gain confidence, and quickly adapt to the technology. Such apps can be conveniently used if they are designed to run on a smartphone or tablet alone (3, 6, 8, 13). The rapid evolution of deep learning techniques has spurred major breakthroughs in automated interpretation of medical images, with notable success reported in dermatological diagnostics and wound evaluation (9). Initial studies utilized standard convolutional neural network architectures, enhanced with domain-tailored adaptations, to tackle the specific computational complexities associated with wound phenotyping (14). The YOLOv5 object detection framework has demonstrated strong computational performance for simultaneously identifying the location of pressure ulcers and grading their severity in clinical photographic datasets (15). The introduction of Vision Transformers has reshaped the foundations of computer vision, offering new possibilities and significant advancements for medical image analysis (16). Existing approaches have largely focused on architectural innovations, yet have given comparatively little attention to the core computational challenges of severe class imbalance and the scarcity of annotated training data that often characterize specialized medical imaging domains (9). Class imbalance remains a major methodological challenge in pressure ulcer classification, as advanced stages (Stage 3 and Stage 4) are substantially underrepresented compared to less severe conditions in clinical datasets (17). Conventional training paradigms tend to generate models with a systematic bias toward majority classes, thereby impairing performance on minority classifications that are of high clinical importance and require prompt intervention (17–20). Recent investigative efforts have explored diverse mitigation strategies including specialized loss function formulations, sophisticated data augmentation protocols, and ensemble learning frameworks to address these distributional limitations (17, 18, 21, 22). Since smartphones are already a common part of daily life, patients and caregivers can use them without having to invest in extra equipment, making healthcare more accessible while saving both time and money. Developing mHealth apps specifically designed to cater for pressure ulcer management offer significant benefits to the patients, physicians, healthcare professionals and caregivers. These apps allow caregivers to confidently care for pressure ulcers in a home setting, while healthcare professionals can remotely monitor the progress through the app. Such innovative apps can improve care and help prevent, assess, monitor, and manage pressure ulcers more effectively. While mobile technology is widely used across the world, the availability of integrated digital solutions for pressure ulcer care, particularly those aimed at home-based caregivers remains limited. This study aims to design and develop a mobile health (mHealth) application named IPI (Interprofessional Pressure Injury), that integrates artificial intelligence-based pressure ulcer staging, caregiver-focused education, personalized nutritional support, and visual wound monitoring features to support caregivers and improve home-based pressure ulcer management. The IPI app derives its name from its development process, which united the contributions of nurses, physicians, dietician, and information and communication technology professionals, all working to ultimately benefit patients and caregivers. Interprofessional collaboration enhances healthcare by bringing together professionals from different fields to deliver more coordinated, consistent, and comprehensive care, ultimately improving both individual patient outcomes and overall public health (23, 24). Materials and methods This study is approved by Institutional Ethics Committee, Kasturba Medical College and Kasturba Hospital (IEC1: 306/2023) and registered under Clinical Trial Registry—India (CTRI/2023/11/059736). The study's data collection for IPI app development began in January 2024, utilizing real patient images to train the model. No images were sourced from public databases or the public domain. All training data were obtained directly from clinical settings at Kasturba Hospital, Manipal, Karnataka, India. A total of 447 images, representing different stages of pressure ulcers, were collected from 342 patients over a 10-months period (January 3, 2024–October 31, 2024) The patient cohort consisted of adults (age range: 18–99 years, mean ± SD: 65.15 ± 20.32 years) receiving inpatient care in the ICU, HDU, and general wards. Images were captured using a Redmi Note 13 5 g (108 MP) camera under ambient lighting with the camera flash. A detailed breakdown of patient demographics and anatomical pressure ulcer locations is provided in Table 1. Informed consents were obtained from all patients before image collection. By using authentic patient data, the development process prioritized both clinical relevance and ethical integrity, laying a strong foundation for accurate and reliable wound assessment in practice. Table 1Summary of patient demographics and dataset characteristics.CharacteristicCategoryValuePatient Demographics (N = 342)Age (Years)Mean ± SD65.15 ± 20.32Range18–9918–3541 (12%)36–64109 (31.9%)65–99192 (56.1%)GenderMale201 (58.8%)Female141 (41.2%)Pressure Ulcer & Image Characteristics (N = 447)Anatomical PositionSacrum184 (41.2%)Ischial167 (37.3%)Elbow32 (7.2%)Ankle28 (6.3%)Heel22 (4.9%)Scapula14 (3.1%)Image Acquisition ProtocolCamera TypeRedmi Note 13 5 g (108 MP)Aspect Ratio4:3LightingAmbient light with flashClinical SettingICU, HDU, general wards Summary of patient demographics and dataset characteristics. A total of 447 images were collected, corresponding to 447 unique pressure ulcers from 342 patients (i.e., one image per ulcer). To establish a reference standard for each image, a rigorous validation process was implemented. Each of the 447 images was independently evaluated by two clinical experts. A critical care nurse practitioner and a critical care physician, both with over 10 years of clinical experience. The experts task was to assign a definitive pressure ulcer stage to each image after it was captured. This process resulted in 100% agreement between the two experts on the final stage for all 447 images. This expert validated stage was then assigned as the image's label for use in model training, and the image file was renamed accordingly. Dataset characteristics and preprocessing We designed a structured framework that integrates advanced data preprocessing methods, a tailored transformer-based model with novel modifications, and comprehensive training and evaluation procedures, all aimed at tackling the specific challenges of classifying pressure ulcers in clinical practice. The study utilized original image dataset of pressure ulcers, gathered directly from clinical settings encompassing six clinically relevant categories such as Stage 1 through Stage 4 pressure ulcers (s1-s4), healthy tissue (he), and undetermined cases (un). The dataset reflected the typical class imbalance often seen in medical imaging, with notably fewer samples in the Stage 3 and Stage 4 categories. These stages correspond to more severe conditions that demand urgent intervention. Class distribution analysis confirmed a highly skewed representation across categories. Interestingly, this distribution aligns with real-world epidemiological observations, where advanced-stage pressure ulcers are less frequent but carry the greatest clinical urgency (17). The inherent imbalance in specialized medical datasets is well recognized as a major challenge in developing robust classification models, particularly for rare but clinically significant pathological conditions (9, 17, 19–22, 25, 26). Prior to processing, all images underwent meticulous quality assessment and standardization. Initial preprocessing involved conversion to a uniform color space, followed by intensity normalization to mitigate variations in illumination conditions across different clinical acquisition settings. Background segmentation was applied to isolate the wound regions, and outlier detection algorithms identified and corrected artifacts from clinical photography, including specular reflections and measurement rulers that could potentially bias the learning process. Advanced augmentation strategy To address the inherent class imbalance and enhance model generalization capabilities, a class-adaptive augmentation framework was developed, dynamically adjusting transformation parameters based on class representation and morphological complexity. All images were initially resized to a standardized resolution of 224 × 224 pixels to ensure dimensional consistency required by the transformer architecture. For underrepresented classes (s3, s4, and un), an aggressive augmentation protocol was implemented including random horizontal and vertical flips (p = 0.7 and p = 0.5, respectively), rotations up to 30 degrees, affine transformations (translation range = 0.2, scale range = 0.8–1.2, shear = 15 degrees), and color jittering with brightness and contrast variation of 0.3. Additionally, we applied perspective distortions (p = 0.3), random auto contrast (p = 0.5), and targeted random erasing (p = 0.3) to simulate partial occlusions and tissue variability. This intensive augmentation strategy was specifically designed to enhance the model's capability to recognize the critical necrotic tissue patterns characteristic of advanced pressure ulcers. For well-represented classes (s1, s2, he), a more conservative augmentation approach was employed with reduced transformation probabilities and ranges. This class-adaptive strategy resulted in a synthetically balanced training distribution while preserving the morphological characteristics essential for accurate staging. All augmented images underwent normalization using the ImageNet statistics [µ = (0.485, 0.456, 0.406), σ = (0.229, 0.224, 0.225)] to facilitate transfer learning from the pre-trained BEiT (Bidirectional Encoder representation from Image Transformers) model. Quantitative analysis demonstrated that our class-adaptive augmentation approach increased the effective sample size of minority classes by a factor of 5.8, while maintaining the clinically relevant visual features necessary for accurate classification. The class-adaptive augmentation approach constitutes an advanced refinement of established medical imaging augmentation methods, purposefully designed to address the distinctive morphological variability and tissue heterogeneity inherent to pressure ulcer pathophysiology (22). The intensive augmentation protocol for minority classes was carefully designed to retain clinically important morphological features while generating enough synthetic samples to balance class distributions across categories (17, 18). This approach has proven especially effective in medical imaging contexts where pronounced class imbalance undermines both model generalization and clinical applicability (18, 22). Enhanced vision transformer architecture The foundation of our classification system is built upon a Vision Transformer architecture, specifically utilizing the BEiT model. We selected the “Microsoft/beit-base-patch16-224-pt22k-ft22k” variant due to its demonstrated efficacy in capturing fine- grained visual patterns through its bidirectional attention mechanisms. The base model was comprehensively extended to address the specific challenges of pressure ulcer classification through several architectural innovations. The choice of BEiT as the base architecture was informed by its demonstrated excellence in extracting fine-grained visual features via sophisticated bidirectional attention mechanisms, capabilities that have consistently delivered strong results in challenging medical image analysis applications (16). The adoption of hierarchical feature representation mitigates fundamental constraints observed in prior transformer-based medical imaging frameworks, wherein single-resolution feature extraction frequently proved insufficient for capturing the multi-scale tissue characteristics and nuanced pathological details essential for accurate analysis (16, 17). The proposed attention mechanism constitutes a targeted refinement of current attention-based strategies in medical image analysis, incorporating relative positional encoding to strengthen spatial relationship modelling within wound regions and enhance discrimination between tissue states of similar morphology. The base model was comprehensively extended to address the specific challenges of pressure ulcer classification through several architectural innovations. Figure 1 illustrates the enhanced BEiT model architecture for pressure ulcer classification. Enhanced BEiT model architecture for pressure ulcer classification, showing the modified transformer blocks, custom self-attention mechanism, and hierarchical feature extraction pathways for improved discrimination between ulcer stages. Hierarchical feature representation The standard BEiT architecture was modified to extract and utilize features at multiple hierarchical levels. Specifically, we integrated outputs from intermediate transformer blocks (6, 9, and 12) through adaptive feature pooling followed by projection layers, allowing the model to simultaneously leverage both low-level tissue texture information and high-level semantic features critical for distinguishing between visually similar ulcer stages. Enhanced self-attention mechanism We developed a specialized attention module that operates on the output CLS token from the BEiT encoder. This module implements an 8-head multi-head attention mechanism with a dimensionality of 768, enabling the model to dynamically focus on discriminative regions within wound images. Importantly, we incorporated relative positional encoding within this attention mechanism to better capture the spatial relationships between different tissue types within the wound bed, a critical factor in accurate staging. The attention module is defined as (27):Attention(Q,K,V)=softmaxx=QKT+Rposdkwhere Rpos represents the relative positional encoding matrix that captures spatial relationships within the wound image. Advanced classification head The classification component was structured as a sophisticated multi-layer perceptron with progressively decreasing dimensions (768 384 6), each block comprising a linear transformation, GELU activation function, layer normalization, and dropout regularization (rate = 0.3). This architecture was specifically designed to form a feature abstraction hierarchy that aligns with the clinical staging criteria for pressure ulcers. Weight initialization followed the Kaiming normal distribution, empirically shown to facilitate gradient flow in deep networks with non-linear activations. The classifier can be formalized as (16):f(x)=W3⋅Dropout(LayerNorm(GELU(W2⋅Dropout(LayerNorm(GELU(W1⋅x))))))The complete architecture comprises approximately 86 million parameters, with the majority concentrated in the transformer encoder blocks. During the initial training phase, we conducted a series of ablation studies to determine the optimal configuration of these architectural components, confirming the significant contribution of each enhancement to the overall classification performance. Training methodology and optimization A rigorous stratified k-fold cross-validation protocol (k = 5) was implemented to ensure robust performance evaluation across dataset variations. The stratification process maintained consistent class distributions across all folds, critical for reliable assessment given the imbalanced nature of the dataset. Each fold comprised approximately 2,000 training images and 400 validation images after augmentation. The class-balanced focal loss is a modified version of the standard focal loss, created to handle the strong class imbalance often seen in pressure ulcer datasets (28). The AdamW optimizer, combined with cosine learning rate scheduling, follows established best practices for training transformer architectures in computer vision. This configuration improves convergence stability and enhances generalization across unseen data (29). Loss function design To address the issue of class imbalance, we implemented a class-balanced focal loss function that incorporates both class frequency and sample difficulty, enabling more effective learning from underrepresented categories (28).LCB=1−β1−βny⋅(1−pt)γ⋅log(pt)where ny represents the number of samples in class y, β is set to 0.9999 to emphasize minority classes, pt is the model's confidence in the correct class, and γ = 3.0 controls the down-weighting of easy examples. This formulation automatically adjusts the learning focus toward underrepresented classes and difficult samples that exhibit subtle discriminative features between ulcer stages. Optimization strategy The optimization process utilized the AdamW algorithm with a finely tuned learning rate of 2 × 10−5 and a weight decay of 0.04 to reduce overfitting. A cosine learning rate schedule with linear warmup during the first three epochs was applied, formally defined as (30):nt=[nmax.tTwarmupift