The machine learning model achieved an AUC of 0.75 with 85.71% sensitivity in detecting chronic stress changes in cardiac patients using wearable sensors.
Can a machine learning model using wearable sensor and chatbot data accurately detect changes in chronic stress levels in post-intervention cardiac patients?
A machine learning model using wearable and chatbot data can reliably detect changes in chronic stress levels in cardiac patients, offering a tool for proactive psychophysiological monitoring.
Absolute Event Rate: 0% vs 0%
Abstract Background Chronic stress is a critical factor influencing cardiovascular outcomes. Elevated stress levels have been linked to poor lifestyle recommendation adherence, increased occurrence of arrhythmias, and increased risk of rehospitalisation or mortality in cardiac populations. Therefore, early identification of psychophysiological vulnerability could enable pre-emptive care in high-risk individuals. Yet, despite growing use of wearables and eHealth tools in cardiology, there is a lack of validated patient-friendly methods to frequently monitor evolution of chronic stress in patients following cardiac interventions. Objective To develop and validate a machine learning model capable of detecting worsening or improvement in chronic stress levels among cardiac patients using longitudinal data from a wearable sensor and a digital health platform. Methods Eighty patients (mean age 73 (17) years, 89% male) used a chatbot-based eHealth platform with wearable sensor for monitoring lifestyle behaviours over 12 months post-intervention. The wrist-worn device collects continuous multimodal data (heart rate, sleep, physical activity, and circadianity) and the chatbot collects subjective data (sleep quality, nutrition intake, and mental stress levels) for seven separate days per month repeatedly during the study period. These data were matched with changes in the Perceived Stress Scale (PSS-10) (a clinically validated questionnaire, our reference method) of which data were collected quarterly. A decrease or increase in PSS-10 score of two points were defined as respectively improved or worsened chronic stress. A tree-bagging ensemble classifier was trained using Bayesian optimization and five-fold cross-validation on a 70/30 training-test split. Class imbalance was corrected via sample weighting. Evaluation metrics included accuracy, sensitivity, specificity, and AUC, with significance tested using the Hanley–McNeil z-test (α = 0.05). Feature contributions were assessed via permutation-based importance. Results The model achieved an AUC of 0.75 (p 0.05), with 78.26% overall accuracy, 85.71% sensitivity, and 66.67% specificity. The confusion matrix indicated reliable differentiation between improvement and worsening in chronic stress levels. Top predictors included daily steps, circadian rhythm (mesor), and age, highlighting both behavioural and physiological markers of stress dynamics. Conclusion This study introduces a novel and interpretable algorithm for monitoring evolution of chronic stress in cardiac patients using an eHealth platform with integrated chatbot and wearable sensor. The validated model provides a foundation for integrating stress profiling into post-cardiac care, enabling more tailored and timely intervention. Such systems address a major gap in current digital health strategies by moving beyond passive tracking to actionable prediction of psychophysiological vulnerability.Figure 1.Classifier performanceFigure 2.Feature importance
Goevaerts et al. (Thu,) reported a other. The machine learning model achieved an AUC of 0.75 with 85.71% sensitivity in detecting chronic stress changes in cardiac patients using wearable sensors.