.png)
Symposia
Suicide and Self-Injury
Brooke Ammerman, Ph.D. (she/her/hers)
Assistant Professor of Psychology
University of Wisconsin-Madison
Madison, WI, United States
Ross Jacobucci, Ph.D.
Research Assistant Professor
University of wisconsin-Madison
Madison, WI, United States
Suicidal ideation (SI) fluctuates dynamically, yet ecological momentary assessment (EMA) often relies on static or semi-random sampling, limiting its ability to capture high-risk periods. An adaptive approach could improve real-time SI detection and inform just-in-time interventions. Leveraging reinforcement learning (RL), we developed the Adaptive Time Interval System (ATIS), a data-driven method that adjusts EMA schedules based on individual risk patterns.
ATIS uses Q-learning, a reinforcement learning algorithm, to tailor assessment timing based on a participant’s recent responses. Predictions are driven by two separate random forest models: one estimating the probability of active SI at the next time point and the other estimating the likelihood of an EMA response. The system selects the next assessment interval (e.g., 30 minutes; 3 hours; next day) based on probabilistic comparisons of Q-values, continuously refining predictions as more data become available.
In a calibration dataset used to train the initial ATIS algorithm (n=53, 7,905 responses), a comprehensive model incorporating lagged SI, time-related factors, and 24 risk factors (e.g., thwarted belongingness, perceived burdensomeness, affect, substance use, NSSI) showed only slight improvement over a simpler model using lagged SI alone (AUC = 0.862 vs. 0.846; AUPRC = 0.491 vs. 0.465). Thus, ATIS prioritizes prior SI responses over additional risk factors.
In a pilot sample (n=14, past-month SI), ATIS accurately predicted whether SI would occur in 30 minutes or 3 hours with an AUC of 0.88, accuracy of 0.87, precision of 0.60, and recall of 0.70.
These findings support the feasibility of using an adaptive system to improve SI monitoring in real-world settings. Data collection is ongoing (target n=40) to further validate the system.
Preliminary findings suggest ATIS enhances EMA scheduling, improving SI risk detection while reducing participant burden. Its ability to adapt to individual patterns highlights RL’s potential for suicide risk monitoring. Future work should explore integrating ATIS into just-in-time adaptive interventions (JITAIs) to deliver personalized support in high-risk moments.