Predicting involuntary hospitalization in psychiatry: A machine learning investigation

Background Coercion in psychiatry is a controversial issue. Identifying its predictors and their interaction using traditional statistical methods is difficult, given the large number of variables involved. The purpose of this study was to use machine-learning (ML) models to identify socio-demographic, clinical and procedural characteristics that predict the use of compulsory admission on a large sample of psychiatric patients. Methods We retrospectively analyzed the routinely collected data of all psychiatric admissions that occurred between 2013 and 2017 in the canton of Vaud, Switzerland (N = 25,584). The main predictors of involuntary hospitalization were identified using two ML algorithms: Classification and Regression Tree (CART) and Random Forests (RFs). Their predictive power was compared with that obtained through traditional logistic regression. Sensitivity analyses were also performed and missing data were imputed through multiple imputation using chain equations. Results The three models achieved similar predictive balanced accuracy, ranging between 68 and 72%. CART showed the lowest predictive power (68%) but the most parsimonious model, allowing to estimate the probability of being involuntarily admitted with only three checks: aggressive behaviors, who referred the patient to hospital and primary diagnosis. The results of CART and RFs on the imputed data were almost identical to those obtained on the original data, confirming the robustness of our models. Conclusions Identifying predictors of coercion is essential to efficiently target the development of professional training, preventive strategies and alternative interventions. ML methodologies could offer new effective tools to achieve this goal, providing accurate but simple models that could be used in clinical practice.

Automated summary

This study used machine learning models to identify predictors of compulsory admission in psychiatric patients, finding aggressive behaviors, referrer, and primary diagnosis as main predictors. Machine learning methods could offer new effective tools to achieve this goal, providing accurate but simple models for clinical practice.