Key result
Wearable-based LSTM model outperforms threshold rules, predicting adverse health events five days ahead with ~94% accuracy.
Why the study?
Chronic NCDs impose substantial mortality and healthcare costs in Malaysia, but current management relies on periodic clinic assessments rather than continuous monitoring.
Does an LSTM-based predictive analysis framework using continuous wearable data improve the prediction of health events in patients with diabetes and/or hypertension compared to other machine learning models?
Does an LSTM-based predictive analysis framework using continuous wearable data improve the prediction of health events in patients with diabetes and/or hypertension compared to other machine learning models?
Absolute Event Rate: 93.7% vs 78.6%
An LSTM-based AI framework using continuous wearable data demonstrated high accuracy in predicting health events up to five days in advance in a synthetic cohort of patients with chronic diseases.
May aid remote monitoring in diabetes and hypertension; hypothesis-generating pending prospective validation.
Chronic Non-Communicable Diseases (NCDs), such as diabetes mellitus, hypertension, and cardiovascular disease, make up a significant portion of the mortality cases in Malaysia, as well as impose consistent demands for costs related to health care expenditure. According to the cost reports at the national level, major NCDs (cardiovascular disease, diabetes, and cancer) are responsible for RM 9.65 billion annually as direct healthcare cost [3]. However, the current approach to managing chronic diseases focuses heavily on periodic assessments in clinic setting at weekly or monthly intervals. Here, a predictive analysis framework based on AI technology, leveraging continuous monitoring with patient data collected from consumer wearables to predict potential health events up to five days ahead is proposed. A synthetic dataset was generated to simulate the behaviour of 500 patients affected by diabetes and/or hypertension, who were continuously monitored throughout a year for heart rate, heart rate variability, physical activity, sleep indices, and blood pressure features. The models are compared to determine the optimal algorithm in the case study. The LSTM model showed 93.7% accuracy in predicting five days in advance, outperforming the random forest model (91.5%), XGBoost model (90.6%), and the simple threshold rule-based approach (78.6%). The sensitivity and specificity scores of the LSTM model were 87.6% and 95.9%, respectively, while the ROC-AUC score was 0.983. According to the SHAP analysis, previous heart rate values (history for three days and seven days) have the largest contributions in predicting patient outcomes, followed by changes in heart rate variability and low activity levels. Although evaluation is based on synthetic data for methodological validation, the framework is intended for extension to real-world wearable streams to support earlier risk detection and more proactive chronic disease management.
No takes yet. Share an insight, caveat, or question.
Razimi et al. (2026) studied Diabetes and/or hypertension (n=500). LSTM predictive model vs. Threshold rule-based approach was evaluated on Accuracy of predicting adverse health events 5 days in advance. An LSTM model using continuous wearable sensor data predicted adverse health events five days in advance with 93.7% accuracy, outperforming a simple threshold rule-based approach (78.6%).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: