An Interpretable Framework Based on Deep Learning and the Internet of Things for Predicting Risk Dynamics in Chronic Patients Publisher

Summary: Can wearable data forecast health risks? An LSTM model reached 95% accuracy in one dataset, with poor sleep and sudden activity shifts prominent predictors. #DigitalHealth #MachineLearning

Pouriayevali B ; Ehteshami A
Source: 2026 International Interdisciplinary Conference on Artificial Intelligence: Engineering, Health, Finance and Humanities, IICAI 2026 Published:2026

Abstract

Introduction: Effective management of chronic conditions necessitates analytical approaches capable of capturing temporal dynamics within physiological data streams. Conventional machine learning methodologies frequently encounter limitations in this domain, functioning as opaque black boxesthat obscure decision-making processes. Objective: This investigation introduces an interpretable framework integrating Internet of Things (IoT) infrastructure with deep learning architectures to forecast health deterioration risks among individuals with chronic conditions. Methods: Temporal data were obtained from the established PhysioNet repository. A long short-term memory (LSTM) network was developed and optimized to anticipate critical health events. Model performance was evaluated against four baseline approaches (artificial neural network and random forest). Furthermore, the SHAP (SHapley Additive exPlanations) methodology was employed to elucidate model predictions and identify predominant risk determinants. Results: The LSTM architecture demonstrated superior predictive capability relative to baseline models, attaining 95% accuracy with an area under the ROC curve (AUC) of 0.97. SHAP-based analysis identified diminished sleep quality and abrupt alterations in physical activity patterns as the most significant predictors of patient risk. Conclusion: The developed framework not only enhances predictive precision but also fosters confidence in and adoption of intelligent health systems through clinically relevant explanatory capabilities. © 2026 IEEE.