Utilizing Machine Learning for Context-Aware Digital Biomarker of Stress in Older Adults

Md Saif Hassan Onim, Himanshu Thapliyal, Elizabeth K. Rhodus

Identifying stress in older adults is a crucial field of research in health and well-being. This allows us to take timely preventive measures that can help save lives. That is why a nonobtrusive way of accurate and precise stress detection is necessary. Researchers have proposed many statistical measurements to associate stress with sensor readings from digital biomarkers. With the recent progress of Artificial Intelligence in the healthcare domain, the application of machine learning is showing promising results in stress detection. Still, the viability of machine learning for digital biomarkers of stress is under-explored. In this work, we first investigate the performance of a supervised machine learning algorithm (Random Forest) with manual feature engineering for stress detection with contextual information. The concentration of salivary cortisol was used as the golden standard here. Our framework categorizes stress into No Stress, Low Stress, and High Stress by analyzing digital biomarkers gathered from wearable sensors. We also provide a thorough knowledge of stress in older adults by combining physiological data obtained from wearable sensors with contextual clues from a stress protocol. Our context-aware machine learning model, using sensor fusion, achieved a macroaverage F-1 score of 0.937 and an accuracy of 92.48% in identifying three stress levels. We further extend our work to get rid of the burden of manual feature engineering. We explore Convolutional Neural Network (CNN)-based feature encoder and cortisol biomarkers to detect stress using contextual information. We provide an in-depth look at the CNN-based feature encoder, which effectively separates useful features from physiological inputs. Both of our proposed frameworks, i.e., Random Forest with engineered features and a Fully Connected Network with CNN-based features validate that the integration of digital biomarkers of stress can provide more insight into the stress response even without any self-reporting or caregiver labels. Our method with sensor fusion shows an accuracy and F-1 score of 83.7797% and 0.7552, respectively, without context and 96.7525% accuracy and 0.9745 F-1 score with context, which also constitutes a 4% increase in accuracy and a 0.04 increase in F-1 score from RF.