AI-enabled remote and objective quantification of stress at scale
Abdulrhman H. Al-Jebrni, Brendan Chwyl, Xiaoyu Wang, Alexander Wong, Bechara J. Saab
Abstract
Accurate measurement of human stress at scale is a major mHealth challenge. Here we explore the potential for deep neural networks (DNNs) to improve remote and objective quantification of stress from voluntary selfie videos captured through mobile device front-facing cameras. Two DNNs were trained with heart rate (HR) and heart rate variability (HRV) data obtained through photophlethysmographic imaging (PPGI) of 11,823 mobile device selfie videos captured in tandem with self-assessments of stress, and compared to contemporary algorithms used to estimate stress from HR and HRV data. A classification DNN and predictive DNN determined self-reported stress with 86 % accuracy and a mean absolute error of 0.001, respectively. Both DNNs performed far better than other recently described approaches when applied to the identical dataset. Well-trained DNNs can objectively and remotely quantify stress at scale. Future efforts may concentrate on the measurement of additional enigmatic cognitive states.