Integrating Heart Rate Variability Improves Machine Learning-based Prediction of Panic Disorder Symptom Severity.

Abstract

Objective: The association between panic disorder (PD) and heart rate variability (HRV) has long been studied with a focus on the imbalance of the autonomic nervous system. This study aims to demonstrate the predictive capability of HRV in determining PD severity using machine learning.

Methods: Psychometric scales and various HRV components were measured from 507 PD patients who were recruited. We designed three experiments with different sets of input features for comparison. The input features of each experiment were 1) both psychometric scales and HRV together (ExSH), or 2) only the scales (ExS), or 3) only the HRV components. In each experiment, nine machine learning models were used to predict the Panic Disorder Severity Scale. We compared the predictive capability of the three sets of input features by statistically analyzing the performance metrics of the models in the three experiments. SHapley Additive exPlanation (SHAP) was further employed to assess the importance of the input features.

Results: The Random Forest model in ExSH, which incorporated both psychometric scales and HRV, achieved the highest f1-score (76.50%) and sensitivity (75.35%). ExSH showed significantly higher sensitivity and f1-score compared to ExS. For the RF model of ExSH, the highest SHAP importance value was found for the Hamilton Rating Scale for Anxiety, followed by the Hamilton Depression Rating Scale, and the low-frequency power (LF).

Conclusion: Our findings demonstrate that integrating HRV with psychometric scales improves machine learning-based prediction of PD severity. We also highlighted LF as a promising variable among HRV components.