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

Abstract

The convergence of smart technologies and predictive modelling in prisons presents an exciting opportunity to revolutionize the monitoring of inmate behaviour, allowing for the early detection of signs of distress and the effective mitigation of suicide risks. While machine learning algorithms have been extensively employed in predicting suicidal behaviour, a critical aspect that has often been overlooked is the interoperability of these models. Most of the work done on model interpretations for suicide predictions often limits itself to feature reduction and highlighting important contributing features only. To address this research gap, we used Anchor explanations for creating human-readable statements based on simple rules, which, to our knowledge, have never been used before for suicide prediction models. We also overcome the limitation of anchor explanations, which create weak rules on high-dimensionality datasets, by first reducing data features with the help of SHapley Additive exPlanations (SHAP). We further reduce data features through anchor interpretations for the final ensemble model of XGBoost and random forest. Our results indicate significant improvement when compared with state-of-the-art models, having an accuracy and precision of 98.6% and 98.9%, respectively. The F1-score for the best suicide ideation model appeared to be 96.7%.

Keywords: Ensemble; Machine learning; Model reduction; SHAP; Smart prisons.

Figure 1. Overview of methodology for classification on the reduced dataset.

Figure 2. Feature contributions by SHAP values excluding suicidal ideation features.

Figure 3. Feature contributions by SHAP values including suicidal ideation features.

Figure 4. Comparing model accuracy, precision and sensitivity with Nordin et al. (2023).

Figure 5. Comparing model F1-score, precision and sensitivity with Horvath et al. (2021).

Figure 6. Comparing model AUC, PPV and log loss with Horvath et al. (2021).

Figure 7. Execution time of our ensemble models on 10-fold cross-validation for performance impact assessment.

Figure 8. Comparing execution time of original and modified anchor library across three runs.

Figure 9. Comparing average execution time of original and modified anchor library for three runs.

Figure 10. Comparing average precision of original and modified anchor library for three runs.

Figure 11. Comparing average precision of original and modified anchor library averaged for three runs.