Personalizing Suicide Risk Assessment: Machine Learning Extraction of Cross-Modal Interactions Between Psychosocial...

Area of Science:

• Computational psychiatry
• Clinical informatics
• Natural Language Processing (NLP)

Background:

• Veterans exhibit elevated suicide risk compared to the general population.
• Current suicide risk prediction models primarily use structured electronic health record (EHR) data.
• Clinical notes offer rich contextual psychosocial information that can be leveraged with NLP to enhance risk detection.

Purpose of the Study:

• To improve suicide risk prediction models for veterans by integrating structured EHR data with unstructured clinical notes.
• To explore the utility of bag-of-words (BoW) representations from clinical notes, outperforming previous semantic approaches.
• To identify and analyze cross-modal interactions between textual and clinical variables for enhanced risk stratification.

Main Methods:

• Utilized a cohort of 27,241 veterans, analyzing clinical documentation from 30, 90, or 270-day lookback windows.
• Employed XGBoost models to balance structured clinical variables with high-dimensional BoW NLP features from clinical notes.
• Evaluated models across different suicide risk strata (low, medium, high) and temporal windows.

Main Results:

• Incorporating NLP-derived interactions into generalized linear models significantly improved predictive performance, especially for low- and medium-risk veterans.
• The BoW representation demonstrated superior performance compared to prior semantic index-based NLP methods.
• A substantial reduction in the performance gap was observed between interpretable and complex predictive models.

Conclusions:

• Interpretable NLP methods effectively uncover clinically meaningful interactions between psychosocial and demographic factors in veteran suicide risk.
• The findings establish a robust benchmark for future deep learning models aiming to capture deeper contextual and temporal information from clinical narratives.
• This approach enhances proactive care strategies for veterans by improving the accuracy and interpretability of suicide risk prediction.