Diabetes is a global health burden, and early detection is critical for timely intervention. This study explores a non-invasive, data-driven framework to identify individuals at risk of diabetes using Volatile Organic Compounds (VOCs) and lifestyle variables. We use causal inference techniques to estimate the impact of VOCs such as acetone, isopropanol, isoprene, and ethanol on blood glucose levels. Additionally, we designed a classifier to distinguish diabetics from non-diabetics using non-invasive markers. We created a risk-based ranking system for individuals in the "gray zone," and identified natural clusters in the population using Gaussian Mixture Model. Our results suggest that specific VOCs exhibit a strong causal influence on glucose levels and that machine learning models can reliably classify and stratify individuals at high risk. This integrated causal-explainable analysis can support the development of tool for non-invasive early screening of diabetes.

Prediabetes is a clinically elusive and often undetected metabolic condition, marking a critical transition from healthy to diabetic states. Non-invasive methods that can identify prediabetes as a statistically and metabolically distinct class remain rare. In this work, we present an AI powered breath analyzing system which is trained on synthetically generated volatile organic compound (VOC) mixtures, that uses a novel knowledge-driven generation technique, to mimic physiologically relevant VOC ranges found in human breath across healthy, prediabetic states. The sensor system generates dynamic response data from these controlled mixtures, which are used to construct machine learning models. The first classification approach is a three-class classification using 210 samples (70 per class), showing improved performance with an F1-score of 0.8 for the prediabetic class using Random Forest and XGBoost. The second approach employs a hierarchical classification model that first detects healthy vs non-healthy breath profiles and then distinguishes between prediabetic and diabetic sub-classes. This two-tier system yields robust identification of non-healthy individuals with 94% accuracy in Tier-1 comprising of both prediabetes and diabetes, and 74% accuracy in Tier-2 for identifying prediabetes despite controlled overlap introduced to capture transitional metabolic states. Our results establish a validated in-vitro foundation for future clinical translation. The proposed framework is now ready for field trials aimed at real-world deployment for affordable and scalable prediabetes screening.

The growing need for real-time privacy-preserving health monitoring has driven the adoption of edge intelligence in biomedical sensing. Alcohol affects cardiovascular health. Detecting alcohol-related cardiac issues enables doctors to identify heart problems early and provide the right treatment. Many patients avoid sharing their drinking habits due to discomfort or social stigma. In this paper, we hypothesize and present a TinyML based system for detecting alcohol-induced abnormalities using raw ECG signals and subject metadata. For this pilot study, we collected raw ECG signals and metadata from hospitalized patients. ECG features were extracted and processed locally, reducing bandwidth and latency while ensuring data privacy. Machine learning models like Random Forest, XGBoost, Logistic Regression, and Linear Support Vector Classifier (SVC) were trained and deployed on Raspberry Pi and Arduino Nano ESP32. Linear SVC achieved the best results with 93% accuracy and F1-scores of 0.89 (Raspberry Pi) and 0.85 (ESP32). This work shows how Edge AI and TinyML can detect alcohol-induced abnormalities from ECG, which is practical, secure, and helps build smarter healthcare systems.

Training medical segmentation networks hinges on our ability to find good minimizer of highly non-convex loss functions, yet the way different losses sculpt this landscape remains poorly understood. Classical objectives such as crossentropy, Dice, and Tversky act as static point wise penalties, offering little insight into how anatomical structure and uncertainty influence optimization. In this work, we reinterpret supervision through the lens of reaction-diffusion dynamics and introduce Morphogenetic Field Loss (MFLoss), which evolves predictions and ground truth as interacting morphogenetic fields before measuring their discrepancy. This view yields smoother, more topology-aware minima that generalize across organs, lesions, layers, instruments, and brain tissue, while remaining fully compatible with standard backpropagation and architectures. Intuitively, diffusion flattens noisy valleys into stable tissue basins, and reaction carves sharp boundary ridges where anatomy demands precision, providing a physically interpretable explanation for the improved convergence and generalization observed in our experiments.

Post-stroke rehabilitation is a critical phase for stroke survivors constituting majority of recovery period. Gamification is a common approach taken towards rehabilitation enhancing patient engagement. This involves using customized and sophisticated sensing setups. In lower-middle income countries (LMICs), acquiring such setups is a challenge, exacerbated by financial constraints post acute-care period and a very low number of therapy centers within practical reach of patients. We are building an accessible & affordable rehabilitation suite targeting such population groups, with an integrated gaming suite as one of the offerings. In this demonstration, we present a 3D game aiming motor function of fingers, which is a critical upper body function towards an independent living. With Google’s Mediapipe running on a smartphone, hand landmarks are detected in real-time and transmitted to the device running the game, which can be a laptop/PC or another smartphone. This makes our solution accessible and can be used by the patients at home in a ubiquitous way without any special setup. Clinically relevant indexes are calculated and shown on screen in an intuitive manner providing sufficient performance insights to both patient and the therapist. The gaming suite is poised for clinical trials along with other assessment modules of our rehabilitation suite, which includes lower-body assessments as well.

Squat is one of the most widely-adopted exercises for strengthening the lower body, especially quadriceps and ham-strings. If not performed correctly, especially as a part of rehabilitation while recovering from a lower body injury, it can cause lower back pain, severe knee injuries, and muscle imbalance. We present a fully automated, 3D posture monitoring and assessment system which works in real-time using Google’s Mediapipe pose land marking, powered with our customized scoring based on the type of squat prescribed by a physiotherapist. Our assessment model calculates clinically relevant parameters like squat depth, squat bias, stance width etc. as defined in literature for a certain prescribed variant of squat, which itself can be recorded as a reference profile in the system by the physiotherapist extracted from sample repetitions performed by them personalized for their patient. A score report is generated for a squat session, quantifying the quality of squat being performed by the patient compared against the prescribed variant. The comparison score is generated on-device using sigmoid-based metric customized for every parameter, without requiring heavy computations or prior dataset like in a DL-based assessment model. The pipeline is robust to environmental and demographic factors. In this demonstration, we aim to show the capability of real-time scoring using multiple reference profiles for different squat variants at the venue. This is a step forward towards a safer exercising and reliable remote rehabilitation monitoring.

predict blood glucose levels using volatile organic compounds (VOCs) detected in human breath. Breath samples from 94 participants, including healthy individuals and those with varying degrees of diabetes, were collected and analyzed using Gas Chromatography-Mass Spectrometry (GC-MS) to quantify VOCs such as acetone, isopropanol, isoprene, and ethanol. These biomarkers were used to develop regression models that predict blood glucose levels non-invasively. The models account for inter-person variability, including factors like smoking habits and stress levels, enhancing their robustness. Our results demonstrate the potential of VOC-based breath analysis as a reliable alternative for glucose monitoring, with future scope for integrating these models into portable nano-sensor systems for real-time diabetes management. Key performance metrics of the model are presented to showcase its predictive accuracy and potential clinical applications. Our best result in predicting blood glucose achieved a 10% MAE, demonstrating the feasibility of using VOCs beyond acetone, with a correlation coefficient (r) of 0.550 and p = 0.023.

Freezing of Gait (FoG) is one of the most debilitating symptoms of Parkinson's Disease (PD), severely impairing mobility and increasing the risk of falls. Traditional FoG detection methods predominantly rely on offline machine learning models, limiting their feasibility for real-time monitoring and wearable deployment. In this study, we present a lightweight and patientinvariant FoG detection framework, specifically optimized for real-time execution on resource-constrained hardware. Our approach utilizes Kullback-Leibler (KL) divergence to measure the similarity between a new patient's gait features and existing cohort data, enabling an adaptive and generalizable detection model. Additionally, we introduce a cohort selection strategy, categorizing subjects into Matched Clusters (MC) and Unmatched Clusters (UC) to enhance model robustness while reducing data dependency. Experimental evaluations on the Daphnet dataset demonstrate that our patient-invariant model achieves comparable classification accuracy, with an average Sensitivity of. 91 and Specificity of. 80, while significantly reducing model size. Furthermore, our approach effectively operates with a single IMU sensor, making it a computationally efficient and practical solution for continuous FoG assessment in real-world applications.

Hair and scalp-related diseases often go unnoticed in the initial stages and patients sometimes cannot differentiate between hair loss and regular hair fall. Diagnosing hair related diseases is time-consuming as it requires professional dermatologists to perform visual and medical tests. There are some works reported in literature which perform binary classification of whether the person has alopecia or not. Often, clinicians calculate SALT, Ludwig or Norwood scores to detect the stage of alopecia. This would require a patient’s scalp view from top, front, sides and back, which acts as a constraint on some datasets that only have a single view of the patient. Our proposed method of scalp detection, skin and hair region segmentation and alopecia detection can be performed on single view image. As the final deployment device is a memory constrained device, our model uses classical image processing algorithms to segment the hair and skin regions. Our method reported an accuracy of 94% on detecting alopecia on the Dermnet dataset. The lack of segmentation ground truth for hair and skin in Dermnet has motivated us to use the Figaro dataset for evaluating the performance of the proposed segmentation methodClinical relevance— Unlike other reported methods in the literature which just states whether the person has alopecia or not, the proposed method provides a complete insight into the percentage of skin and hair regions on the scalp by performing segmentation of skin and hair regions along with classification. The report generated from the method help the dermatologists to monitor the efficacy of the treatment by comparing varied dated images before and during the treatment cycle. The proposed method can be run on portable devices even with limited memory for monitoring the successfulness of alopecia treatment at home.

Blood pressure (BP) serves as a critical indicator of cardiovascular health. Hypertension, a grave medical condition, manifests when the body's BP surpasses the threshold of 140/90 mmHg, leading to life-threatening risks such as heart attacks and strokes. Many commercial digital BP devices based on Oscillometric method are often less accurate which may lead to inaccurate diagnosis, especially in home settings. This research attempts to develop an accurate BP prediction method using deep learning model over korotkoff sound and cuff pressure. This is a pilot study to validate the method of BP prediction and to show proof of concept. A digital BP measurement prototype device was developed on the principle of auscultation, renowned as the 'gold standard' in manual BP measurement, and it is used for the sole purpose of capturing excellent quality korotkoff sound and cuff pressure. Using this prototype device, data was collected from a cohort of over 115 subjects, with measurements conducted under the supervision of medical professionals. A Convolutional Neural Network (CNN) based model was trained on this captured data to discern Korotkoff peaks, enabling an exact prediction of corresponding systolic and diastolic BP values. After comparing the predictions with ground truth, the validation results show mean absolute error and standard deviation error for systolic and diastolic errors as 4.06 +/- 5.35 mmHg and 3.89 +/- 4.51 mmHg, respectively. This promising outcome underscores the potential of accurate BP prediction using auscultatory method and deep learning as a valuable tool.

Stroke patients face impaired motor and cognitive functions affecting quality of life. Early rehabilitation and clinical assessments can assist in recovery, with home-rehabilitation a convenient approach for the patient. But, it usually employs specialized setups and sensing hardware, impacting its adoptability, especially in low to medium income countries (LMICs). To address this, we present a framework for telerehab using off-the-shelf commercial devices providing novel analytics for motor assessment, and interactive gamified exercises. We present our approach of directly using 3-D swing trajectories of foot calculated from a single 6-axis motion sensor attached to shoe’s sole hitherto possible only with more sophisticated sensing modalities, and using it to assess gait quality using a novel comparative spatial scoring approach. We also present Range of Motion (ROM) assessment which can be performed with a smartwatch or a strapped smartphone providing accurate angle and speed profile, while also catering to stray rotations owing to muscle weakness in stroke patients. An initial proof of concept is presented for both solutions using motion data from smartwatch and comparing spatial outputs with Kinect V2. We also introduce two games from our gaming suite which receive real-time motion input from limb-worn device, aimed at improving patient’s motor and cognitive abilities, with customized scoring methodologies for personalized therapy and evaluating progress of recovery. In future, these games, along with our proposed novel analytics for gait and ROM, are planned for validation on stroke patients in clinical trials.Clinical relevance — Stroke is one of the top contributors of global DALY (Disability-Adjusted Life Years), with 90% contribution from LMICs. This is primarily due to lack of any rehabilitation received post-hospitalization owing to inaccessibility and cost of rehab and its aids, on top of cost incurred on clinical visits. With home-based rehabilitation using easily accessible devices providing rich analytics, not only can there be a reduction in the clinical burden, but also post-stroke care can become more adoptable, affordable and effective.

Hip arthroplasty is a surgical procedure that involves the replacement of a patient’s hip joint with a prosthetic implant. While these implants are initially effective, they may eventually fail and necessitate revision surgery. It is important to identify the 3 Charnley and 7 Gruen zones around the implant and then identify the zone-wise radiolucency which indicates loosening for effective pre- and post-operative planning. Despite the importance of zones, there is a lack of automation attempts in this field. In this work, we have proposed a 3-stage algorithm that detects the sanity of the image for diagnosis, then segments the implant regions into the zones, and then identifies radiolucency within the zones. We have demonstrated a 94% accuracy for Fit/Not Fit segregation, a 0.95 dice score for our zonal segmentation, and a 98% overall loosening accuracy. Obtaining an average dice score of 0.92 in the segmentation of zones and 0.93 accuracy on loosening detection on a blind dataset indicates the robustness of the proposed algorithm.Clinical relevanceThe detection of loosening of Joint replacement is an indicator of the prosthesis failing and needing a revision. The extent of loosening along the zones helps to determine the level of difficulty and the type of implant that would need to be kept ready for revision arthroplasty. Our work provides these critical information to the surgeons and helps them towards better preplanning of revision surgery.

Multi-modal image segmentation has significant potential for advancing high-quality representation learning, as different modalities provide complementary information about anatomical structures, organs, and diseases. However, there is currently no principled approach to designing compact, edge-efficient architectures that effectively leverage multi-modal images for medical image analysis. Existing practices are either manual, relying on expert-driven fusion of features from unregistered and unpaired modalities, or employ overly large architectures with multiple encoders and decoders.In this paper, we bridge this gap by introducing Y-Net, a novel architecture tailored to jointly learn segmentation tasks for multiple organs using data from diverse medical imaging modalities. Y-Net, combined with an automated hyperparameter search methodology, can be deployed in a plug-and-play fashion on multi-modal imaging datasets, delivering accurate and parameter-efficient organ segmentation. We validate Y-Net on the CHAOS challenge task, which involves segmenting abdominal organs from CT and MRI data. Our approach achieves a 6% improvement in Intersection over Union (IoU) scores across all organ classes compared to state-of-the-art single-modality segmentation methods, while requiring only one-twentieth of the parameters. This compactness makes Y-Net particularly well-suited for on-premises, privacy-preserving inference in healthcare analytics.

Biological Age (BA) indicates the authentic ageing progression of an individual in relation to their quality of life. The noninvasive identification of BA is crucial in predicting longevity and early age-related diseases and enabling personalized healthcare. Potential biomarkers of BA are vague and need attention. The ageing process stands out as a prominent risk factor for cardiovascular diseases. Consequently, an Electrocardiogram (ECG), the most popular and easily accessible signal, is explored to analyze the effect of age. Numerous studies have delved into supervised deep-learning approaches for ECG analysis, particularly in predicting age. These studies rely on regression-based methods and necessitate additional analysis for extracting health-related insights, such as the correlation of error between Chronological Age and AI-predicted Age with mortality. Moreover, as the shortage of cardiologists’ annotated data is apparent, we propose an Age-Informed Convolutional Autoencoder that clusters ECG deep features associated with age to assess the quality of life possessed at the current age. We also proposed a three-step training strategy combining model training and deep ECG features clustering with a controlled initialization. We find that a combination of age and ECG reveals the heart’s BA and is a contributing biomarker for estimating the overall BA of the body. This approach marks substantial progress in analyzing age-related impacts on ECG. It provides new perspectives on different cardiovascular disorders and can potentially transform personalized healthcare in the future.

As the global population ages, chronic diseases such as Parkinson's disease (PD) are on the rise, placing significant strain on healthcare systems. This paper presents a conceptual framework for automatic personalized assistive systems-integrating wearable sensors, mobile apps, and machine learning-to enhance PD management. These systems aim to deliver real-time interventions and tailored treatment plans to address the complex symptoms of PD. Successfully addressing challenges such as user acceptance and technical reliability could improve patients' quality of life and alleviate healthcare burdens.

The rapid advancements in medical technology throughout the past two decades have greatly enhanced life expectancy, ushering in an era where ageing populations are thriving. However, this demographic shift has also brought attention to age-related issues, particularly neurological disorders. This paper addresses the challenges posed by one such neurological disorder Parkinson’s disease (PD), with a focus on the debilitating symptom of Freezing of Gait (FoG). We propose a novel algorithm utilizing wrist-worn accelerometer sensors for real-time FoG detection. The algorithm incorporates Variable Mode Decomposition (VMD) to enhance noise resilience in motion data. At its core, our emphasis has been on developing a subject-independent model and harnessing the capabilities of a wrist-based sensor, enhancing the feasibility and usability of our approach. Evaluation on a public dataset showcases the model’s significant potential, with an achieved average Sensitivity of 0.80 and Specificity of 0.81, underlining its effectiveness.

We present a smart-phone RGB camera based system for automatic human 3D posture monitoring driven by our proposed Machine Learning (ML) backbone. Rather than mapping RGB image sequences directly to 3D posture, we learn an intermediate, concise feature space representation of 3D postures. To learn this feature space, we train an auto-encoder learnt on 3D posture sequences. We train a mapping from the image sequence to this concise 3D feature space. This mapping subtracts all unnecessary information present in the image domain and retains only 3D posture information. This elimination helps in dataset independence and increases robustness of our model. The mapped feature space vector is then mapped to 3D posture domain. For estimating 3D human pose, most States-Of-the-Art need either a 3D depth sensor which is expensive or need multiple cameras for stereo-vision. Our proposed cross-modal method uses monocular video from single low-end smart-phone camera for estimating human 3D pose in real time (similar to 5 fps) with high accuracy (less than 1 cm error per joint location). We also show our method being effective for posture analysis in yoga and gym exercises.

Exhaled breath contains hundreds of volatile organic compounds (VOCs) linked to various biochemical pathways, hence tracking their concentration levels useful for the early detection, substaging, and prognosis of health conditions. Elevated levels of acetone and isoprene, which are by-products from metabolic pathways of lipids, are correlated with specific medical disorders. This study demonstrates the potential of an indium-oxide-based sensor to detect trace amounts of these VOCs in human breath at parts-per-billion (ppb) levels in an artificial breath-like environment. A selectivity analysis was conducted to differentiate the sensor's response to the targeted VOCs from other gases typically present in breath, mitigating the risk of false positives. The sensor demonstrated stable and reproducible performance over ten weeks, validating its practicality for real-world use. In addition, this article explores various machine learning (ML) techniques to enhance VOC analysis and proposes an end-to-end framework for the automatic classification and prediction of VOC concentrations. Peak-based feature extraction is used to effectively profile VOCs, and Gaussian-noise-based data augmentation is used to improve the accuracy of classification and regression models. This comprehensive approach integrates advanced sensor technology with ML methods to offer a robust solution using noninvasive breath analysis for early detection and follow-up of diseases related to raised lipid metabolism rate.