Suicide Death Prediction Using the Maryland Suicide Data Warehouse: A Sensitivity Analysis

Area of Science:

• Public Health
• Data Science
• Clinical Informatics

Background:

• Suicide causes nearly 50,000 deaths annually in the US.
• Contextualizing suicide risk models is crucial for effective clinical decision support.
• Existing models require evaluation based on their intended use and generalizability.

Purpose of the Study:

• To assess the performance of different predictive models for suicide risk.
• To evaluate how model results can inform clinical decision support systems.
• To analyze the generalizability and utility of suicide risk models.

Main Methods:

• Utilized the Maryland Suicide Data Warehouse (MSDW) for a 4-year retrospective study.
• Compared binary logistic regression with ridge and LASSO penalized regression models.
• Employed fivefold cross-validation and evaluated models using sensitivity, positive predictive value (PPV), and F1 score.

Main Results:

• Male sex, depressive/anxiety disorders, social needs, and prior attempts were associated with suicide death.
• Cross-validated binary logistic regression outperformed penalized models.
• Models achieved low-to-moderate PPV and sensitivity, with a peak F1 score of 0.323.

Conclusions:

• Suicide death prediction is context-dependent, requiring a balance between precision and recall.
• Model evaluation should align with the specific level of clinical intervention.
• The utility of predictive models may vary across different healthcare settings and needs.