Development of a machine learning model for predicting compulsory psychiatric care using clinical notes

BACKGROUND: If patients at high risk of compulsory psychiatric care can be identified sooner, it is in some cases possible to apply stage-specific interventions that make compulsory care unnecessary. We therefore developed a machine-learning algorithm that could identify patients at risk sooner, thereby enabling timely intervention and preventive measures.

METHODS: As predictors in a machine-learning model for predicting compulsory care, we used known risk factors and clinical notes registered in electronic health records during treatment at a mental health institution in the Netherlands.

RESULTS: We successfully built a predictive model that, with a precision of 0.91, identified just over 50% of the compulsory care trajectories (recall/sensitivity) 60 days before the start of the trajectory. The algorithm was thus able to predict more than half of the compulsory-care admissions 2 months before they took place. It correctly flagged nine out of ten cases as being at high risk for admission. A previous compulsory-care episode was the most predictive variable in our model. Incorporating text improved the model's performance, particularly when the forecast horizon was shorter.

CONCLUSIONS: This study shows that predicting adverse events is a promising line of research that may support future clinical practice. To enhance the clinical relevance of predictions, it is essential to understand the functioning of the model and its input variables.