Predicting inmate suicidal behavior with an interpretable ensemble machine learning approach in smart prisons.

Area of Science:

• Artificial Intelligence
• Criminology
• Public Health

Background:

• Smart technologies and predictive modeling offer potential for inmate behavior monitoring and suicide risk mitigation.
• Existing machine learning models for suicide prediction often lack interoperability and detailed interpretability.
• Current interpretation methods focus on feature importance, overlooking rule-based explanations.

Purpose of the Study:

• To develop an interpretable machine learning framework for predicting inmate suicide risk.
• To introduce Anchor explanations for generating human-readable rules in suicide prediction models.
• To enhance model performance by combining SHAP and Anchor explanations with ensemble methods.

Main Methods:

• Utilized SHapley Additive exPlanations (SHAP) for initial feature reduction in high-dimensionality datasets.
• Employed Anchor explanations to create simple, human-readable rules for model interpretation.
• Developed an ensemble model combining XGBoost and random forest, refined by SHAP and Anchor interpretations.

Main Results:

• Achieved significant improvements over state-of-the-art models in suicide risk prediction.
• Attained an accuracy of 98.6% and a precision of 98.9%.
• The best suicide ideation model demonstrated an F1-score of 96.7%.

Conclusions:

• The novel approach enhances the interpretability and accuracy of suicide risk prediction models in correctional settings.
• Combining SHAP and Anchor explanations offers a robust method for understanding complex predictive models.
• This research paves the way for more effective, technology-driven suicide prevention strategies in prisons.