Advancing Diabetes Detection with Denoising Autoencoders and Softmax in a Generative Deep Learning Pipeline

Main Article Content

Swapnil P. Suryavanshi
Nitin B. Pawar

Abstract

We propose a state-of-the-art generative AI architecture for hypoglycemia and hyperglycemia prediction and avoidance that integrates autoencoders for deep physiological feature extraction with a Softmax classifier for glucose-state classification. Instead of neuroimaging, the system processes continuous glucose monitoring (CGM) data along with a comprehensive set of patient-specific physiological, behavioral, and lifestyle parameters, including age, weight, insulin dosage (basal and bolus), duration of medication usage, physical activity intensity, stress level, dietary intake patterns, and sleep quality. This multimodal data provides a deeper understanding of glucose fluctuations and individual metabolic responses. The autoencoder efficiently encodes these high-dimensional CGM signals and associated parameters into compact latent representations, reducing noise while preserving clinically relevant glucose-dynamics patterns. These distilled features are then passed to a Softmax classifier that accurately distinguishes between normal glucose levels, impending hypoglycemia, and impending hyperglycemia. The proposed method leverages neural-networkbased modeling of glucose variability to enhance both prediction accuracy and computational efficiency. Experimental evaluation on benchmark CGM datasets demonstrates strong performance, achieving a prediction accuracy of 93.75%, highlighting the capability of the framework to support early detection of dangerous glucose deviations. By providing timely warnings for both low and high blood glucose events, this system has the potential to significantly improve proactive intervention, reduce risk, and support personalized diabetes management.

Article Details

Section
Articles