ARES-SI: Adaptive Reinforcement-Enhanced Sampling for Suicidal Ideation

Suicidal ideation (SI) fluctuates rapidly within individuals, with approximately 50% of its variance occurring within-person. Standard ecological momentary assessment (EMA) protocols use fixed schedules that fail to capture these rapid fluctuations, missing critical risk windows and opportunities for intervention. Here, we present ARES-SI (Adaptive Reinforcement-Enhanced Sampling for Suicidal Ideation), a reinforcement learning system that personalizes EMA timing based on individual risk patterns. ARES-SI combines random forest models to predict active SI probability and non-response likelihood with a Q-learning algorithm to determine optimal assessment intervals (30 minutes for high-risk states, 3 hours for low-risk states, and next day to reduce burden). We trained the system using EMA data from 98 participants across three studies with varied sampling frequencies (4–15 assessments per day; 2,359 positive active SI responses) over 28-day protocols. In validation testing (N = 36; 353 positive active SI responses), the random forest models achieved an AUC of 0.865 for active SI prediction and 0.750 for non-response prediction. ARES-SI demonstrated 65% higher action selection accuracy than random scheduling (0.781 vs. 0.474; macro F1 = 0.650 vs. 0.377), with superior detection of high-risk windows (recall = 0.810 for 30-minute assessments vs. 0.392 for random) and efficient low-risk scheduling (precision = 0.972 for 3-hour assessments vs. 0.876 for random). These findings demonstrate that reinforcement learning can identify when individuals are most vulnerable, providing the temporal precision needed for just-in-time adaptive interventions in suicide prevention.