Personalised machine-learning decision support for suicidal thoughts and behaviours in the psychiatric emergency...

Individuals presenting with psychiatric complaints to the emergency department (ED) are at significant risk of re-referral for suicidal thoughts and behaviours (STB) in the short term. The absence of effective tools to identify these high-risk individuals hampers clinical decision-making, necessitating the development of improved, interpretable, and personalised prediction methods. We aimed to predict the risk of suicide ideation (SI) and suicide attempt (SA) within one, six, and 12 months using XGBoost algorithms; and enhance model interpretability through methods in explainable artificial intelligence (XAI).

This prognostic study conducted the initial development and evaluation of machine learning (ML) models using structured electronic health records (EHR) from 26,198 patients (51,397 referrals) at the psychiatric emergency department (PED) of the University Hospitals Leuven, Belgium, from January 1, 2002, to December 31, 2022.

The models achieved good discriminative ability (Area Under the Curve) ranging from 0·70 [95 % CI: 0·66 to 0·75] to 0·80 [95 % CI: 0·75 to 0·85] and excellent calibration, with Integrated Calibration Index values ranging from 0·001 to 0·004 suggesting reliable predictions. The referrals identified within the five highest-risk ventiles comprised 24·73 - 26·19 % of all referrals, which accounted for 50·43 % -70 % of all STB referrals. The models exceeded the minimum criteria for cost-effective targeting of active contact and follow-up interventions. Applying XAI methods improved the transparency and interpretability of the predictions on a patient level.

Enhancing ML models based on PED EHRs with methods in local XAI has the potential to aid clinicians in effectively identifying high-risk individuals for targeted, cost-effective interventions.