Digital biomarker discovery of emotional stress periods through change point detection and supervised machine learning

Music City Center 

In this talk, we introduce a combined framework that detects emotionally stressful periods in older patients with chronic pain and depression and establishes data-driven rules that explain the onset of their stress periods. Despite the well-established link between stress exposure and the worsening of emotional distress and mood disorders in these patients, data-driven and long-term intervention algorithms that function without human supervision remain lacking. The first part of the framework focuses on detecting emotional stress periods, referred to as hotspots. We propose a change point detection algorithm that tracks changes in localized, low-level statistical features of passively sensed data and self-reported stress levels. The second part of the framework focuses on identifying which statistical features in which passive sensing variables are most critical in triggering the hotspots. We leverage the signed iterative random forest (siRF) to derive machine-derived rules that can serve as guidelines for determining whether patients are in a stress period at a future time. We present simulation studies under various scenarios and showcase applications of the algorithms to ALACRITY Phases I and II data. The proposed framework is expected to help form digital biomarkers from passively sensed data, enabling personalized and data-driven interventions.

Mental health

mHealth

change point detection

variable selection

random forest

high-order interaction