Temporally informed random forests for suicide risk prediction

Area of Science:

• Computational psychiatry
• Machine learning in healthcare
• Public health informatics

Background:

• Suicide is a leading global cause of death, posing a challenge for accurate risk identification by clinicians.
• Existing algorithmic suicide risk detection models, often using electronic health records (EHRs), have limitations in leveraging temporal data for improved predictions.

Purpose of the Study:

• To develop and validate a novel temporally enhanced random forest (RF) model for suicide risk detection.
• To assess the impact of incorporating temporal information into RF models for predicting suicide risk.

Main Methods:

• Proposed the Omni-Temporal Balanced Random Forests (OT-BRFs) model, integrating temporal information into each tree.
• Utilized longitudinal EHRs and clinician notes from Mass General Brigham Health System (1998-2018).
• Compared OT-BRFs against a Naive Bayes Classifier and standard balanced RF models.

Main Results:

• Temporal variables, such as visit frequency and EHR coverage duration, were associated with suicide risk.
• RF models outperformed Naive Bayes classifiers in predicting suicide risk (AUC 0.824 vs. 0.754).
• The OT-BRF model demonstrated superior performance, achieving 0.339 sensitivity at 95% specificity, outperforming other RF models.

Conclusions:

• Temporal variables play a crucial role in suicide risk detection.
• Incorporating temporal information across all RF trees enhances predictive performance.
• Integrating temporal context into risk prediction models improves the interpretation of patient data and overall accuracy.