Age-related macular degeneration (AMD) is one of the leading causes of vision impairment among older adults. Deep learning (DL) techniques have shown significant potential for accurate AMD detection with the help of optical coherence tomography (OCT) images. This paper proposes a deep learning framework that integrates image preprocessing, retinal layer segmentation, feature identification, and classification for automated AMD detection. Initially, OCT images are enhanced using a hybrid filtering approach to suppress noise while preserving structural information. The enhanced images are then segmented using a transformer-based network with attention and multi-scale feature fusion to capture local and global retinal information. Next, complementary texture and shape features are extracted and fused into a unified feature representation. Finally, the fused features are classified using a residual learning network integrated with AlexNet and an enhanced channel attention mechanism to improve discriminative feature learning. The experimental findings indicating that the proposed framework outperforms existing methods, achieving a 98.1% detection accuracy using 90% of the dataset, demonstrating its effectiveness and robustness for automated AMD diagnosis.
Sports injuries are common among athletes, often leading to extended recovery and compromised performance. Early and accurate detection of injuries enables timely intervention and improved rehabilitation. However, existing solutions such as wearable devices, computer vision systems, and clinical assessments are limited by complex deployment requirements, user discomfort, or the need for expert supervision.We present GoodVibes, a non-intrusive injury detection system that passively captures gait cycles from footstep-induced floor vibrations on sports facility walkways. Designed for seamless integration into athletic training environments, this system monitors natural walking and captures gait measurements for personalized injury detection. The two key challenges in developing GoodVibes are: (1) overlapping injury signatures caused by musculoskeletal complexity, and (2) inter-person variability reflected in both vibration patterns and personal training or injury history. To address these challenges, GoodVibes introduces three complementary solutions: (1) Variance-Guided Hierarchical Classification, a hierarchical divide-and-conquer approach to reduce uncertainty caused by musculoskeletal interdependence, where adjacent muscles compensate for injured regions, producing overlapping vibration signatures; (2) Dual Gait Vectors, a personalized representation that contrasts reference and observed gait to isolate individual-specific deviations and mitigate uncertainty arising from inter-person variability; and (3) Context-Aware Learning, which integrates personal injury history and specialized sports categories as contextual embeddings to further reduce prediction uncertainty.We evaluated GoodVibes in lab and real-world sports stadium deployments with 21 student-athletes. The system achieved 92.8% accuracy in laboratory settings and 84.2% accuracy in real-world deployments, demonstrating strong performance and potential for long-term passive monitoring in athletic environments.
Inertial sensors have become significant tools for assessing human movement due to their ability to provide precise data on acceleration, velocity, and orientation in a non-invasive manner. This bibliometric analysis examines the evolving landscape of scientific literature on the use of Inertial Measurement Units (IMUs) in sports over the past two decades. Quantitative methods were employed to analyze 12,620 documents retrieved from the Scopus and Web of Science databases, aiming to identify research trends, key authors, and influential publications. A significant and rapid growth in scientific production appeared, with a strong and growing citation impact (h-index = 192; +27% of papers above the field average), led by the USA and China. Thematic analysis identified “biomechanics” as a basic theme, “wearable sensors” and “machine learning” as core themes, and “gait”, “balance”, and “physical activity” as more developed topics. “Sensor fusion”, “activity recognition”, and “deep learning” emerge as the most impactful themes in terms of citations. In contrast, high-frequency terms such as “sensor” and “wearable” exhibit a lower return. A shift towards deep learning and human activity recognition confirms the growing integration of artificial intelligence in IMU-based movement pattern recognition, performance analysis, injury prevention, and rehabilitation monitoring.
Psychological stress is a major risk factor for several physical and mental health conditions, motivating the development of reliable and unobtrusive monitoring solutions. Wearable devices enable continuous and ecological acquisition of physiological and physical signals, providing a promising basis for automatic stress detection in real-world settings. In this context, machine learning (ML) and deep learning (DL) methods have shown strong potential for extracting stress-related patterns from multimodal wearable data. However, most studies evaluate performance on a single dataset, limiting insights into model robustness and generalization across populations and experimental conditions. Furthermore, there is currently no comprehensive comparison between feature-based ML methods and data-driven DL methods. Finally, the use of transfer learning to address differences across various datasets has not yet been thoroughly explored.In this work, we investigate automatic stress detection from smartwatch data by comparing feature- and data-driven models, focusing on cross-dataset generalization. Experiments were conducted on four public datasets collected with the same wearable device and sensing modalities, including photoplethysmography, acceleration, skin temperature, and electrodermal activity. We first evaluated model performance within each dataset to establish baseline results. We then assessed generalization using cross-dataset testing, where models trained on one dataset are evaluated on unseen datasets. Finally, we explored different transfer learning strategies to improve cross-dataset performance and promote generalization, especially in small datasets.Results show that DL methods (F1-score 0.69–0.96) consistently outperform ML models (F1-score 0.69–0.95) in within-dataset evaluations, although performance varies across datasets. Cross-dataset tests reveal significant performance degradation, particularly for DL models, where the F1-score showed an average drop of −21%. Transfer learning mitigates this gap, improving robustness across datasets. These findings highlight the importance of cross-dataset evaluation for realistic assessment of wearable-based stress detection systems.
This study assessed the usability, satisfaction, and acceptability of a 5G-enabled telemedicine service among Internally Displaced Persons (IDPs) in Abuja, Nigeria. A quantitative pre–post design was employed, with baseline and endline data collected from 75 participants who completed telemedicine consultations. Structured Likert-scale questionnaires captured demographics, access to mobile devices, and telemedicine outcomes. Data were analyzed using paired-samples t-tests, independent-samples t-tests, one-way ANOVA, and Pearson correlations. Participants were predominantly female (69.3%) and aged 40–49 years (29.3%). Most had secondary education (36.0%), and 60% had lived in the camp for over two years. Only 2.7% reported regular electricity access, while 21.3% owned a device suitable for telemedicine consultations. Mean usability scores increased from 77.16 (SD = 10.68) at baseline to 83.83 (SD = 10.86) at endline (t = 3.85, p < .001). Satisfaction improved from 42.07 (SD = 9.35) to 47.53 (SD = 8.56) (t = 3.60, p = .001), and acceptability rose from 39.08 (SD = 6.12) to 43.03 (SD = 5.19) (t = 4.05, p < .001). While improvements were consistent across age groups, variations were observed by gender, device ownership, and electricity access, with greater gains among males and participants with better access. Strong positive correlations were observed between usability and satisfaction (r = 0.751), usability and acceptability (r = 0.849), and satisfaction and acceptability (r = 0.797). Overall, a nurse- and translator-supported teleconsultation service delivered over a 5G network was feasible and was associated with short-term positive changes in self-reported usability, satisfaction, and acceptability among a self-selected group of IDPs in this resource-constrained setting, despite limited infrastructure and access barriers; in the absence of a control group, causal attribution of these improvements to 5G connectivity specifically is not possible from this design.
Human Action Recognition (HAR) has greatly improved with the integration of spatial, structural, and kinematic sensor modalities. However, real-world implementation on resource-constrained edge devices is severely limited by the high computational cost of deriving dense kinematic trajectories (optical flow). Moreover, combining these diverse streams often leads to topological incompatibilities and modality imbalance. This research proposes a unique resource-aware Tri-Stream HAR framework to tackle these obstacles. First, we process 2D skeleton coordinates using a dimensionally aligned Hierarchical Temporal Convolution Network (HT-ConvNet) to resolve topological constraints. We present Compressive Normalized Fusion to ensure optimal stability during multi-sensor integration. By actively avoiding high-variance visual features from controlling the structural data, this technique balances variance across streams. We eliminate computational redundancy by designing a dynamic routing pipeline that accounts for uncertainty. A lightweight bi-modal “Fast Pass” that uses only spatial and structural data is used to start the inference process. By calculating the Shannon Entropy of this initial prediction, the system measures predictive uncertainty in real time. It only initiates the computationally expensive kinematic extraction for highly ambiguous sequences. Extensive evaluations across the JHMDB benchmark and an NTU RGB+D 60 subset validate this architecture. The framework achieved a mean accuracy of 83.12% on JHMDB, with a peak split accuracy of 87.06%. More importantly, the dynamic entropy gate routed up to 92.8% of sequences strictly through the Fast Pass. This selectively bypassed optical flow extraction, yielding substantial computational savings, reducing total system GFLOPs by 59.2% to 84.1% without degrading predictive power. Ultimately, this framework provides an accurate, numerically stable, and computationally efficient solution for real-world multimodal HAR.
Massage therapy is a popular technique that helps individuals reduce stress, pain, and muscle tension. However, the effectiveness of massage therapy is compromised by different factors that can cause unintentional injuries such as incorrect force level during the massage and areas where massage is not directly targeting the pain. To address these issues, the development of wearable massage monitoring system has been proposed to improve the efficiency and safety of massage therapy and training. In this work, we developed fabric-based force sensor array and communication protocol via ZigBee platform which are integrated to form a wearable real-time massage monitoring system. The force intensity, trigger points, and duration of massage in specific areas are gathered to evaluate the effectiveness of the massage therapy and training. The results demonstrate that the system can monitor a broad range of force measurements (1 to 250 N). The response and recovery times are less than 0.1 s. Movement massage steps can be classified into four stages; (I) pause stage (II) emphasis stage (III) stationary stage, and (IV) lifting stage according to the force levels during the massage. These stages are suitable for use in both massage therapy and training. The process of constructing the wearable massage monitoring system is systematically proposed step by step.
Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, and continuous real-time monitoring outside clinical settings is still limited. However, traditional cardiovascular tests are usually performed only during hospital visits and often miss short-lived or developing abnormalities, which makes continuous monitoring an important clinical need. The proposed architecture combines contrastive pretraining, a convolutional neural network, a bidirectional long short-term memory (BiLSTM) network, an attention mechanism, and focal loss-driven finetuning to address data imbalance and weak supervision. Electrocardiogram (ECG), photoplethysmogram (PPG), and respiration signals from the BIDMC PPG and Respiration dataset are used to train and evaluate the model. The framework achieves 96.66% accuracy and a 0.963 ROC-AUC, while additional subject-independent evaluation provides a more realistic assessment of generalization to unseen individuals. Grad-CAM, t-SNE, and permutation-based modality analyses support clinical interpretability by highlighting discriminative waveform regions and the relative contribution of each modality. The inference latency is approximately 25 ms per 48-s window, demonstrating feasibility for edge deployment. Overall, the proposed framework supports predictive and preventive cardiovascular monitoring and personalized care in both clinical and remote P7 healthcare settings, with potential integration into wearable devices, body sensor networks, and ICU monitoring systems.
Specialized sensors are included in smart wearable devices engineered to provide effective human fall detection and continuous monitoring of routine activities. These devices utilize mainly gyroscopes and tri-axial accelerometers to record the patterns of body movement. The collected sensor data are processed by signal preprocessing pipelines to improve detection accuracy. The pipelines include noise reduction, feature extraction, and data normalization. In this study, a comprehensive survey of wearable fall detection systems is presented, focusing on the sensor, data processing methods, and fall detection techniques. This paper compares traditional threshold-based methods and state-of-the-art (SOTA) Machine Learning (ML) models, such as Convolutional Neural Networks (CNNs), Long Short-term Memory (LSTM) networks, and SOTA transformer models, in terms of detection performance, computational efficiency, and applicability to resource-limited wearable devices. It also highlights the importance of wireless communication technologies such as Bluetooth and Wi-Fi to send emergency notifications to caregivers or health care providers. But there are many problems, such as high energy consumption, delayed response, false alerts from normal daily activities, user variation, and an unbalanced data set. Moreover, this research focuses on emerging research trends and technical challenges in the design of real-time, reliable, and energy-efficient fall detection systems for the elderly and vulnerable populations. The outcomes of this study offer valuable insights into the efficiency of current approaches, shed light on limitations in current research, and propose future directions to improve wearable fall detection systems.
Background The rapid emergence and spread of infectious diseases challenge traditional surveillance methods. Artificial intelligence (AI) offers a modern solution, yet its integration and performance in disease surveillance across sub-Saharan Africa (SSA) remain uneven, with wide variation in infrastructure and adoption. This review maps AI applications in infectious disease surveillance in the region, focusing on prediction, detection, and response. Methods A comprehensive literature search was conducted for studies published between 1 January 2015 and 15 August 2025 across PubMed, EBSCOhost, Web of Science, and Scopus. The search followed the PCC framework, PRISMA guidelines, and Arksey and O'Malley's methodology. Two independent reviewers screened articles with inter-rater agreement assessed, and a third reviewer resolved conflicts. Methodological quality was appraised using the Mixed Methods Appraisal Tool (MMAT), and findings were synthesized narratively using thematic analysis. Results Out of 3143 articles, 62 were included, covering 13 countries and more than eight infectious diseases, mostly malaria and COVID-19. Most studies were descriptive quantitative in rural, urban, or mixed settings, with 60 (97%) rated high quality. Machine learning (ML) methods were most frequently used, followed by deep learning (DL) and hybrid/ensemble models, with k-fold cross-validation being the predominant validation strategy. Five themes emerged: predictive modelling, spatial/geospatial surveillance, informing public health decision-making, disease classification, and diagnostic enhancement. Conclusion This review shows that AI can enhance infectious disease surveillance in SSA, though applications remain limited to a few diseases, highlighting the need for broader coverage, standardized datasets, cross-country and external validation, and context-specific ethical frameworks.
Physiological signal-based prediction systems for allergic rhinitis (AR) exacerbation hold substantial promise for preventive medicine. However, existing approaches predominantly rely on medical-grade sensors in controlled laboratory settings, leaving their feasibility for real-world deployment with consumer wearables largely unvalidated. This study presents a systematic comparative analysis of AR exacerbation prediction across medical-grade and consumer wearable sensors, investigating whether appropriately designed deep learning (DL) architectures can compensate for inherent signal quality limitations in consumer devices. We acquired synchronized physiological signals—heart rate variability (HRV), electrodermal activity (EDA), and skin temperature—from AR patients and healthy controls during controlled thermal exposure using parallel medical-grade and consumer-grade sensor configurations. To address cross-device signal heterogeneity, we implemented robust feature-extraction pipelines that employ Welch's spectral estimation for HRV frequency-domain metrics and cvxEDA decomposition for EDA components. We subsequently developed device-specific DL architectures optimized through systematic Bayesian hyperparameter tuning, explicitly integrating AR medical history as augmented clinical features. Leave-One-Subject-Out (LOSO) cross-validation across 10 repeated training runs demonstrated that our optimized DL framework achieved F1-scores of 0.89 ± 0.01 and 0.82 ± 0.01 for comfortable and discomfort states, respectively, in a binary classification framework using medical-grade sensors. Critically, consumer-grade sensors achieved comparable performance (0.86 ± 0.01 and 0.75 ± 0.01)—markedly superior to baseline artificial neural networks (0.83 ± 0.01 and 0.68 ± 0.02)—while simultaneously reducing memory footprint from 1.30 MB to 1.07 MB through clinical feature integration. This narrow performance gap demonstrates that appropriately designed DL models can effectively bridge the limitations of consumer sensors, demonstrating the feasibility of consumer-grade wearables as a scalable alternative for AR exacerbation monitoring in preventive medicine.
The 12-lead electrocardiogram (ECG) is the gold standard for cardiovascular monitoring, offering superior diagnostic granularity and specificity compared to photoplethysmography (PPG). However, existing 12-lead ECG systems rely on cumbersome multi-electrode setups, limiting sustained monitoring in ambulatory settings, while current PPG-based methods fail to reconstruct multi-lead ECG due to the lack of explicit inter-lead constraints and insufficient modeling of spatial-temporal dependencies across leads. To bridge this gap, we introduce P2Es, a physiology-conditioned diffusion framework that integrates several synergistic innovations. First, to resolve the physiological information gap, we propose GroupFinder, a contrastive alignment module that retrieves optimal affinity matrices by aligning the PPG latent space with ECG-derived clusters. Furthermore, to mimic real-world distortions, the forward process applies frequency blurring and temporal noise, while the reverse process employs a hierarchical multi-scale generation module with frequency deblurring to recover fine-grained diagnostic details. Finally, we implement a resource-efficient edge architecture for rapid on-device inference. Extensive experiments demonstrate that P2Es significantly outperforms state-of-the-art baselines. This provides a scalable solution for longitudinal cardiac screening and pre-clinical risk stratification.
Respiratory diseases, including COVID-19, influenza, MERS-CoV, pertussis, and the common cold, remain significant public health concerns, particularly in indoor environments such as schools and universities. A major challenge in infectious disease modeling is the estimation of transmission rates between individuals. Previous studies have largely focused on population-level or spatiotemporally averaged transmission measures, leaving individual contributions to disease spread insufficiently characterized. To address this gap, we deployed an Ultra-Wideband (UWB)-based Real-Time Location System (RTLS) to collect high-resolution positional data during four social gatherings on a University campus. These data were used to perform contact analysis and to estimate individual-level transmission rates and reproduction numbers for multiple respiratory diseases. By moving beyond aggregated estimates to person-specific transmission metrics, this framework captures heterogeneity in contact behavior and provides a more realistic representation of indoor transmission dynamics. Such individual-level characterization has important implications for improving infectious disease modeling and strengthening targeted public health surveillance in shared indoor spaces. To facilitate adoption and reproducibility, we further developed an open-source Python package that enables researchers and public health practitioners to estimate individual transmission metrics from indoor tracking data and apply these methods to the analysis of respiratory disease spread in real-world settings.
Remote photoplethysmography (rPPG) has emerged as a promising technology for non-contact heart rate estimation from facial videos. However, most existing methods produce deterministic estimates without reporting reliability which is a critical limitation for real-world deployment. We present PhysQual, a framework that treats trustworthiness as a first-class output by jointly estimating blood volume pulse (BVP) signals and multi-dimensional quality scores including motion, illumination, skin visibility, and temporal consistency available at inference time. We introduce three complementary strategies: (1) Physics-Guided Quality by estimating reliability from environmental dynamics; (2) Performance-Guided Quality by learning a proxy to identify prediction errors due to latent failures, and (3) Uncertainty-Guided Quality by capturing input-dependent uncertainty through learned variance. Experiments on UBFC-rPPG and TokyoTech datasets demonstrate that PhysQual significantly improves accuracy and reliability awareness. Our framework achieves a 52% error reduction in similar-domain tests (2.68 -*1.28 bpm) and 16% in cross-domain scenarios (3.26 -* 2.73 bpm), outperforming established baseline models. Further, error-retention analysis shows that PhysQual enables tunable reliability at inference time, reducing cross-domain error by 34.6% through targeted rejection of unreliable samples. These results highlight PhysQual's flexibility for robust rPPG deployment across diverse operating scenarios.
Emergency chest X-ray (CXR) triage has a structural modality gap: reports arrive after triage decisions, yet multimodal foundation models require image-text inputs. We present Variational Risk Minimization (VRM), a distillation framework that treats LVLM-generated report variants as Monte Carlo samples of latent clinical interpretations. Rather than distilling from a single teacher target, VRM learns from a variationally marginalized teacher distribution, enabling uncertainty-aware supervision under missing-modality constraints. Under matched encoder families, VRM outperforms direct fine-tuning baselines and improves calibration with strong recovery from hallucinated supervision. Marginalized supervision reduces report-selection instability. In our compact edge-student instantiation, a confidence-gated cascade reaches AUC 0.941 at 103ms average latency with 20.3% cloud escalation, yielding an explicit reliability-latency operating point for cloud-edge clinical workflows.
Vaccination significantly contributes to healthy aging; however, older adults frequently face cognitive, accessibility, and communication barriers that result in suboptimal vaccine uptake. This systematic literature review identifies key determinants of vaccine hesitancy and evaluates technological strategies to improve immunization rates among older adults. We conducted a structured search of the PubMed database for studies published between 2020 and 2025, with the final search completed on 20 December 2025. Using a custom 1–3 relevance and quality scoring framework, findings from 27 eligible studies were synthesized. Our analysis reveals a complex landscape of cognitive, accessibility, and language-related barriers, alongside promising interventions such as personalized messaging, targeted letters, and automated reminders. Evidence demonstrates that tailored communication delivered via user-centered mHealth applications can effectively reduce vaccine hesitancy and improve uptake. Personalized outreach aligned with individual health profiles, communication preferences, and timing emerges as a key enabler of sustained engagement. Building on these insights, this review examines how context-aware mHealth systems can be designed to promote vaccination among diverse aging populations. We propose GoldenVaxShield, a context-aware, AI-powered mHealth solution that leverages personalized messaging, reminders, and real-time feedback to support informed decision-making and timely vaccination. By integrating advanced technologies with personalization and accessibility principles, GoldenVaxShield aims to enhance preventive healthcare engagement, improve vaccination coverage, and support healthy aging. While variability in study methodologies and a focus on technology-driven interventions remain considerations, the findings highlight the potential of personalized digital outreach as a scalable pathway to improving population-level vaccine uptake.
Background: Digital twin technology, which creates virtual replicas of physical entities through real-time data integration and computational modelling, has emerged as a transformative innovation in healthcare. Its application in obstetrics and gynaecology represents a paradigm shift in personalised medicine, predictive analytics, and clinical decision support systems. Aim: This narrative review aimed to comprehensively evaluate the current applications, benefits, challenges, and future prospects of digital twin technology in obstetrics and gynaecology, while identifying gaps in existing literature and proposing directions for future research. Methods: A comprehensive literature search was conducted across PubMed, Scopus, Web of Science, IEEE Xplore, and Google Scholar databases from database inception through June 2025. Relevant peer-reviewed articles, case studies, and technical reports were included. Evidence was categorised according to a four-tier maturity framework ranging from theoretical concepts to clinical translation. Results: Digital twin applications in obstetrics encompass fetal growth monitoring, placental function assessment, preterm birth prediction, and labour simulation. In gynaecology, applications include endometrial cancer modelling, ovarian function simulation, menstrual cycle prediction, and surgical planning for complex procedures. The majority of identified evidence currently remains at the theoretical framework or in silico model development stage, with largescale clinical validation trials representing a significant gap. Evidence maturity is more advanced in obstetrics than in gynaecology. The technology demonstrates potential for enhancing diagnostic accuracy, treatment personalisation, risk stratification, and patient outcomes while reducing healthcare costs. Conclusion: Digital twin technology holds substantial promise for revolutionising obstetric and gynaecological care through data-driven personalised medicine. However, implementation challenges including data privacy concerns, computational complexity, validation requirements, and regulatory frameworks must be addressed before widespread clinical adoption. Specific attention must be given to algorithmic bias, the digital divide, and the distinction between technical feasibility and clinical implementability.
Classification of medical ultrasound images is challenging due to heterogeneous imaging conditions, device variability, and subtle tissue features. We present a framework combining physics-informed neural networks (PINNs) with physics-consistent data augmentation to improve accuracy, robustness, and generalization in ultrasound image analysis. Experiments were conducted on a multi-domain liver fibrosis dataset (6343 images; five and three-class tasks) and a merged breast cancer dataset (1719 images; benign, malignant, normal). Images were standardized to 224 & times; 224 pixels in grayscale. Augmentation strategies included classical image augmentation (CIA) for geometric variability and physics-informed augmentation (PIA) to simulate ultrasound-specific artifacts. Three CNN backbones (VGG16, EfficientNetB0, ResNet50) were trained across six regimes that combined augmentation and PINN loss. Results show that PINN regularization consistently improves accuracy, F1-score, and convergence speed. Liver fibrosis classification reached 0.98 accuracy with a 64% reduction in epochs, while breast ultrasound accuracy increased by over 32% relative to baseline, achieving 0.90 with PIA. T-SNE analyses reveal more compact, discriminative feature embeddings. This pipeline showed that physics-informed learning effectively improves clinical ultrasound image classification.
