Enhancing Suicide Attempt Risk Prediction Models with Temporal Clinical Note Features.

Area of Science:

• Medical Informatics
• Clinical Data Science
• Computational Psychiatry

Background:

• Accurate suicide attempt risk prediction is crucial for timely intervention.
• Structured electronic health record (EHR) data alone has limitations in capturing complex patient risk factors.
• Integrating unstructured clinical notes can provide richer insights into patient risk.

Purpose of the Study:

• To enhance a structured-data-based suicide attempt risk prediction model using temporal Concept Unique Identifiers (CUIs) from clinical notes.
• To evaluate the impact of different temporalization schemes and model types on predictive performance.
• To determine the value of unstructured clinical data in improving suicidality risk prediction.

Main Methods:

• Utilized diagnostic codes for suicide attempts within 30, 90, or 365 days.
• Compared models using only structured EHR data (medications, diagnoses, demographics) against hybrid models incorporating unstructured text data.
• Employed fixed 90-day windows and flexible epochs for temporalizing clinical notes.
• Trained and assessed random forests and hybrid long short-term memory (LSTM) neural networks.

Main Results:

• Models incorporating temporal CUIs from clinical notes outperformed structured-data-only models.
• The window-temporalized LSTM model achieved the highest Area Under the Precision Recall Curve (AUPRC) for 30-day predictions.
• Hybrid models generally demonstrated improved performance across various metrics compared to control models.

Conclusions:

• Incorporating EHR-derived clinical note features significantly enhances suicide attempt risk prediction models.
• Unstructured clinical data plays a critical role in improving the accuracy of suicidality prediction.
• Future research should explore advanced methods to further refine prediction accuracy and intervention effectiveness.