Deep Reinforcement Learning-Enabled IoT Framework for Real-Time and Personalized Diabetes Diagnosis Using Wearable Sensors

Abstract

The increasing number of diabetes patients worldwide reveals an immediate necessity for technical diagnostic solutions that offer smart operations and full data protection capabilities. This research presents FL-Hybrid, which unites the CNN-LSTM-Attention network structure together with Dueling Deep Q-Network (DQN) to perform individual diabetes screening using wearable sensor information through federated learning. The PhysioNet BIG IDEAs dataset with CGM and physiological sensor signals feeds into the model that combines spatiotemporal transformers with context-driven decision functions. The system protects user privacy through its deployment on edge devices, combined with federated learning training that eliminates the need to send sensitive information to centralized servers. The CNN-LSTM backbone finds deep physiological patterns, and the attention mechanism focuses on crucial diagnostic indicators such as post-meal glucose spikes and sleep-related heart rate variability decreases. The Dueling DQN agent develops optimum diagnostic actions through integration of clinical accuracy with efficiency and prediction reliability measures. Experimental tests conducted on multiple models show that FL-Hybrid outperforms all other systems by reaching a 97.85% accuracy rate with 97.31% precision, 97.96% recall, and 98.41% AUC-ROC. These results exceed those of CNNs and RNNs as well as centralized DRL models. The model displays 95.34% accuracy when noise occurs while also demonstrating exceptional user adaptation ability at a 95.62% adaptation rate. The proposed system provides organizations with a scalable, privacy-focused solution for continuous diabetes monitoring, which represents an advanced mobile health technology.

Received: 30 May 2025 | Revised: 3 December 2025 | Accepted: 28 January 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support the findings of this study are openly available in PhysioNet at https://doi.org/10.13026/zthx-5212, reference number [29].

Author Contribution Statement

M. Alamelu: Conceptualization, Methodology, Software, Validation, Writing – review & editing, Visualization. Meera Alphy: Software, Validation, Formal analysis, Data curation, Writing – original draft. Finney Daniel Shadrach: Conceptualization, Methodology, Writing – original draft, Writing – review & editing, Visualization, Supervision. Jayaraj Velusamy: Data curation, Writing – review & editing, Project administration.