GAN-Augmented LSTM Framework for Smartphone-Based Human Activity Recognition and Early-Stage Diabetes Risk Assessment

This research contributes to the field in several key ways. First, it introduces a novel application of Human Activity Recognition (HAR) extending it beyond conventional uses like step-counting and fall detection into early diagnostic screening for Type-2 Diabetes by correlating daily activity patterns with known diabetic symptomatology. Second, it addresses the persistent data-scarcity challenge in deep learning-based HAR by demonstrating that GAN-generated synthetic sensor data can meaningfully augment limited real-world datasets, improving LSTM classification accuracy (98.48% vs. 97.79% on real data alone) offering a reusable strategy for HAR researchers facing similar data constraints. Third, the framework achieves reliable recognition at a low 1 Hz sampling frequency, in contrast to the 50 Hz typically used in prior work, making continuous 60-day passive monitoring on smartphones significantly more battery-efficient and practically deployable. Fourth, rather than a binary outcome, the proposed cosine-similarity-based risk score provides a graded, clinically interpretable risk estimate, which was validated against an independent A1C biomarker (6.1%), lending external clinical credibility to the approach. Collectively, this positions the work as a low-cost, non-invasive pre-screening tool with potential to identify at-risk individuals particularly relevant given that a large proportion of diabetes cases worldwide go undiagnosed while also laying groundwork for future real-time, activity- and diet-aware insulin-dosage recommendation systems.