Timing Is Everything: Modeling the Temporal Dynamics of Suicidal Ideation
Abstract
Objective: Suicidal ideation (SI) is a critical clinical concern, yet most predictive approaches rely on static or aggregate measures that overlook its session-to-session dynamics. This study aimed to improve next-session SI prediction by integrating dynamic item-level data and identifying the clinical features that contribute most strongly throughout treatment. Methods: Data from 2,045 outpatients were analyzed across the first 10 sessions. A transformer-based time-series model was trained to predict next-session SI using prior-session item-level scores and was compared against a static elastic-net baseline. SHAP values were calculated to quantify feature contributions at both group and individual levels. Results: The dynamic model significantly outperformed the static baseline (average R² = .95 vs. .12), and performance remained improved even without autoregressive inputs (average R² = .24). While prior-session SI was consistently the strongest predictor, depressive affect, hopelessness, meaning in life, and anxiety-related symptoms also provided significant predictive value. SHAP analyses further revealed temporal fluctuations and patient-specific patterns in influential features. Conclusion: Dynamic, interpretable models using item-level data substantially enhance short-term SI prediction. These findings provide clinical insights into which symptoms are most relevant for specific patients at different times, supporting more responsive and personalized suicide risk monitoring.
Related
- Functional connectivity of anterior cingulate cortex subregions in suicidal ideation in major depression: circuit mechanisms, exploratory classification, and treatment-related changes.
- A comparative analysis of machine learning, deep learning, and ensemble learning models for suicidal ideation detection on social media
- AI-based detection of suicidal ideation in text: model development and evaluation for a student mental health chatbot
- Association Between Suicidal Ideation and Negative Affect: 6-Month Ecological Momentary Assessment Study
- Multi-temporal UAV multispectral phenotyping and machine learning for modeling cotton defoliation dynamics across diverse accessions
BibTeX
@misc{loray2026timing,
title = {{Timing Is Everything: Modeling the Temporal Dynamics of Suicidal Ideation}},
author = {Juan Segundo Peña Loray and Julian Rubel and Juan Martín Gómez Penedo and Miriam Hehlmann and Tobias Steinbrenner and Jay Lawrynowicz and Wolfgang Lutz},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/vk4pw_v2},
url = {https://osf.io/preprints/psyarxiv/vk4pw_v2/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Timing Is Everything: Modeling the Temporal Dynamics of Suicidal Ideation
AU - Juan Segundo Peña Loray
AU - Julian Rubel
AU - Juan Martín Gómez Penedo
AU - Miriam Hehlmann
AU - Tobias Steinbrenner
AU - Jay Lawrynowicz
AU - Wolfgang Lutz
PY - 2026
DA - 2026/09/17
PB - PsyArXiv
DO - 10.31234/osf.io/vk4pw_v2
UR - https://osf.io/preprints/psyarxiv/vk4pw_v2/
AB - Objective: Suicidal ideation (SI) is a critical clinical concern, yet most predictive approaches rely on static or aggregate measures that overlook its session-to-session dynamics. This study aimed to improve next-session SI prediction by integrating dynamic item-level data and identifying the clinical features that contribute most strongly throughout treatment. Methods: Data from 2,045 outpatients were analyzed across the first 10 sessions. A transformer-based time-series model was trained to predict next-session SI using prior-session item-level scores and was compared against a static elastic-net baseline. SHAP values were calculated to quantify feature contributions at both group and individual levels. Results: The dynamic model significantly outperformed the static baseline (average R² = .95 vs. .12), and performance remained improved even without autoregressive inputs (average R² = .24). While prior-session SI was consistently the strongest predictor, depressive affect, hopelessness, meaning in life, and anxiety-related symptoms also provided significant predictive value. SHAP analyses further revealed temporal fluctuations and patient-specific patterns in influential features. Conclusion: Dynamic, interpretable models using item-level data substantially enhance short-term SI prediction. These findings provide clinical insights into which symptoms are most relevant for specific patients at different times, supporting more responsive and personalized suicide risk monitoring.
ER -