Objective evaluation of 35 heart rate variability parameters for predicting takeover performance in conditionally automated driving

Abstract

Drivers will be free to engage in Non-Driving Related Tasks (NDRTs) during Level 3 conditionally automated driving. However, drivers may not be able to respond to takeover requests quickly and flawlessly if the level of mental workload invested in the NDRTs is non-optimal. Heart Rate Variability (HRV) has been reported to be a sensitive indicator of the mental workload during NDRT engagement, and some HRV parameters have been used in Machine Learning models that attempt to predict takeover performance. However, until now, the selection of HRV parameters has been ad hoc. The present study constructed an artificial intelligence model to conduct an unbiased evaluation of 35 HRV parameters for predicting takeover performance in various contexts. The model used performance data from 19 drivers collected under 3 NDRTs x 2 time intervals (6 conditions) in a driving simulator. The HRV parameters were ranked by the predictiveness of takeover performance, using two ground truths. The optimal data ranges for the nine most influential HRV parameters were identified, enabling the optimal level of mental workload during NDRT engagement to be inferred. The present study introduced four innovations: 1) the unbiased use of all HRV parameters; 2) treating each NDRT as a separate condition; 3) using SHAP analysis to identify the most influential parameters; 4) using SHAP analysis to identify the range of values for a given parameter that are associated with an optimal TOR. While previous researchers have focused on time-domain HRV parameters, this study demonstrated that frequency-domain and non-linear parameters offer comparable predictive power. Our novel approach optimises HRV parameters in an unbiased manner, enhancing the prediction of driver takeover performance and improving the development of driver monitoring and warning systems.