Towards Adaptive Stress Reduction Interventions with LLMS: A Baseline Study

• Jacob Gordon, Computer Science, University of Delaware

• Matthew Mauriello, Computer Science, University of Delaware

Digital mental health interventions (DMHIs) for stress reduction traditionally rely on reinforcement learning algorithms, such as multi-armed bandits, to optimize intervention timing and content. However, intervention adherence and efficacy are heavily dependent on user context and emotional state, highlighting a need for deeper personalization. Recent advancements in Large Language Models (LLMs) provide a novel approach to tailoring intervention language; to bridge this gap, we introduce EmotionStream, an LLM-powered, personalized digital stress reduction system. To rigorously evaluate EmotionStream, we must first establish a validated stress-induction protocol to simulate high-stress conditions. Our target population comprises Computer Science (CS) students, a demographic historically exhibiting stress levels significantly higher than both general adults and other undergraduates. In this ongoing preliminary pilot study with expected N=10 participants, we designed a stress-induction task where participants completed increasingly difficult, Leetcode-style competitive programming questions under strict time constraints. Stress levels were assessed using a 5-point self-report scale, passively sensed facial action units, gaze, smartwatch-based heart rate, direct observation of physical stress indicators, a post-task survey, and an exit interview. By confirming that these curated programming tasks successfully provoke substantial stress, this pilot study will validate the stress-induction protocol necessary for our forthcoming, large-scale efficacy trials of the EmotionStream application.