Machine learning detecting aggression among mood disorder patients visited in psychiatric emergency departments

Aggression, which is highly prevalent in patients with mood disorders, has been proven valuable in detecting the progression from hypomania to manic episode, enabling a timely diagnosis and treatment.

This study aims to develop a machine learning model and explore variables detecting and distinguishing aggression among patients visited psychiatric emergency departments.

This cross-sectional study utilized machine learning in detecting aggressive behaviors in patients with mood disorders who attended the emergency department of Beijing Anding Hospital. Four machine learning models-logistic regression with LASSO, random forest, gradient-boosting tree, and decision tree-were developed for detecting aggressive behaviors. These machine learning included clinical variables, biochemical variables (C-reactive protein (CRP), cortisol, Adreno-Cortico-Tropic-Hormone (ACTH), testosterone), demographic variables, and psychometric data, including scores from the Hamilton Depression Rating Scale (HAMD), Hamilton Anxiety Rating Scale (HAMA), Young Mania Rating Scale (YMRS), and Brief Psychiatric Rating Scale (BPRS) for detecting aggressive behaviors. Shapley Additive Explanations (SHAP) analysis was used to interpret the important features.

The random forest model achieved the highest detective accuracy with an area under the curve (AUC) of 0.79 (95 % CI: 0.73-0.85) followed by gradient-boosting tree, 0.77 (95 %CI: 0.71-0.83). Key detector included YMRS and BPRS total scores, while HAMD and HAMA scores identified as protective factors. For depressive episode patients, the random forest achieved the highest detective accuracy with an area under the 0.73 (95 %CI: 0.64-0.83). And for manic episode patients, gradient-boosting tree achieved the highest detective accuracy with an area under the 0.89 (95 %CI: 0.86-0.92). The SHAP analysis highlighted the prominence of manic and psychiatric symptom severity in aggression detection CONCLUSIONS: These findings underscore the potential of machine learning in enhancing risk assessment for aggression in patients with mood disorders, though model calibration requires further refinement. This study demonstrates the transformative potential of machine learning, particularly the random forest model in detecting aggression among patients with mood disorders. These findings highlight the power of data-driven approaches to enhance psychiatric risk assessment, enabling precise, personalized interventions in emergency settings.