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

Comprehensive Summary

This study, conducted by Gericke et al., investigates the efficacy of XGBoost machine learning in predicting suicide ideation (SI) and attempt (SA) among recurrent emergency department (ED) patients with psychiatric complaints. The study data were derived from electronic health records (26,198 patients, at the University Hospitals Leuven, Belgium, over a span of 20 years (January 1st, 2002 - December 31st, 2022). From the results of this study, it was found that the average number of referrals per patient was 1.96, and SI and SA combined constituted the main reason for 24.79% of all referrals. Patients were also more likely to be re-referred for SA or SI if they were referred at first for SA or SI, respectively. The machine learning models were trained to predict either SI or SA re-referral within 30 days, 6 months, or 12 months. For SI prediction, machine learning sensitivity scores for predictions within 30 days, 6 months, and 12 months were 0.82, 0.81, and 0.83, while specificity scores for predictions within 30 days, 6 months, and 12 months were 0.51, 0.50, and 0.51, respectively. For SA prediction, the sensitivity scores for predictions within 30 days, 6 months, and 12 months were 0.90, 0.85, and 0.86, while specificity scores for predictions within 30 days, 6 months, and 12 months were 0.51, 0.53, and 0.50, respectively. From this, it was concluded that machine learning models trained on electronic health records were able to effectively predict suicide re-referral within 30 days, 6 months, and 12 months of initial ED referral.

Outcomes and Implications

Suicidal thoughts and behaviors are present in many mental disorders, such as substance use disorder and major depressive disorder, and among those who seek treatment for SA or SI, a significant portion of patients experience more than 1 referral within a year for SA or SI. However, from previous studies, it has been found that due to factors such as the large scope of the affecting variables and time constraints, suicide risk, as determined solely by physicians, is not highly accurate. Utilizing machine learning like in Gericke et al. can potentially reduce the error in suicide risk assessment by reducing the resources necessary to analyze large amounts of patient data.