Deep Sense: Deep Learning for Early Staging the Onset of Diabetes Emanating Recognition of Activity Patterns

This study presents a comprehensive framework for Human Activity Recognition (HAR) using smartphone sensor data, with a specific focus on identifying activities associated with diabetic risk factors. We investigate the efficacy of two deep learning architectures (Long Short-Term Memory (LSTM) networks and Graph Convolutional Networks (GCNs)) for activity recognition under conditions of limited training data. To address data scarcity, we employ Generative Adversarial Networks (GANs) to augment the training dataset with synthetically generated sensor data, enhancing model robustness beyond what is achievable with real sensor data alone. Continuous accelerometer and gyroscope data spanning daily activities were collected from experimental subjects over a 60-day period. This dataset was used to train and evaluate both LSTM and GCN models, with results demonstrating that the GCN architecture achieves superior performance in recognizing diabetes-related activities such as sedentary behavior, physical inactivity, and irregular meal patterns. Furthermore, we propose a novel risk quantification method that estimates diabetes risk by analyzing the duration and frequency of engagement in diabetes-related activities. We employ cosine similarity to measure the correspondence between activity patterns of diagnosed diabetic patients and experimental subjects, yielding a quantitative risk score. To validate the proposed framework, we conducted clinical HbA1c (A1C) assays on experimental subjects. One subject exhibited an A1C level of 6.1%, corresponding to a prediabetes diagnosis, which corroborated the high-risk classification predicted by our framework. These results demonstrate that the proposed HAR-based approach can accurately assess diabetes risk and classify individuals according to clinically validated diagnostic criteria, offering potential applications in continuous health monitoring and early intervention strategies.