Importance of variables from different time frames for predicting self-harm using health system data.

Area of Science:

• Biomedical Informatics
• Clinical Informatics
• Health Data Science

Background:

• Self-harm risk prediction models often use historical patient data.
• Data availability varies across different time periods prior to an index visit.
• Understanding the temporal importance of predictors is key for model implementation.

Purpose of the Study:

• To assess the predictive potential of variables from different time horizons for self-harm risk.
• To apply algorithm-agnostic variable importance techniques in a biomedical informatics context.
• To inform the implementation of self-harm risk prediction models considering data limitations.

Main Methods:

• Utilized variable importance to quantify predictor potential for self-harm risk.
• Analyzed recent (≤3 months) and distant (>1 year) mental health information from seven health systems.
• Defined predictiveness using area under the receiver operating characteristic curve (AUC), sensitivity, and positive predictive value.

Main Results:

• Mental health predictors from the three months preceding the index visit demonstrated significant importance.
• Excluding recent predictors decreased AUC from 0.85 to 0.77 in one health system.
• Predictors from more distant time frames showed lower importance.

Conclusions:

• Recent predictors are highly important for self-harm risk prediction.
• Implementation challenges arise when recent data is unavailable due to processing lags.
• Variable importance analysis guides the practical application of risk prediction models in clinical settings.