TITLE:
Machine Learning Prediction of Aggression Risk in Psychiatric Patients: A Multi-Modal Approach Integrating Clinical History, Behavioural Patterns, and Physiological Signals
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Aggression Prediction, Machine Learning, Psychiatric Inpatients, Physiological Biomarkers, Heart Rate Variability, Violence Risk Assessment, Precision Psychiatry, Behavioural Monitoring, Artificial Intelligence, Clinical Decision Support
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.5,
May
29,
2026
ABSTRACT: Aggressive behaviour in psychiatric inpatient settings represents a significant clinical and safety challenge, with approximately 30% of patients experiencing at least one aggressive episode during hospitalization. Current risk assessment relies primarily on clinical intuition and static rating scales, lacking predictive validity for time-sensitive intervention. We developed and validated a multi-modal machine learning framework integrating clinical history, behavioural patterns, and physiological signals to predict aggression risk within a 7-day window. We conducted a retrospective analysis of 2500 psychiatric inpatients across multiple diagnostic categories including schizophrenia, bipolar disorder, post-traumatic stress disorder, and major depressive disorder. The study cohort comprised adults aged 18 - 85 years admitted to acute psychiatric units. We extracted 25 features across four domains: demographic characteristics, clinical history (prior aggression, diagnosis, hospitalization patterns), behavioural indicators (irritability, sleep disturbance, medication adherence, substance use, social withdrawal), and physiological biomarkers (heart rate variability, skin conductance, cortisol levels, body temperature deviation). Four machine learning algorithms were evaluated: Random Forest, Gradient Boosting, Logistic Regression, and Support Vector Machine. Model performance was assessed using area under the receiver operating characteristic curve (AUC-ROC), F1-score, and cross-validation. Feature importance was analysed using SHAP values and partial dependence plots. The Random Forest classifier achieved the highest predictive performance with an AUC-ROC of 0.84 (95% CI: 0.81 - 0.87) and F1-score of 0.78. Cross-validation confirmed robustness with mean AUC of 0.82 (±0.04). The most predictive features were prior aggression count (mean |SHAP| = 0.142), current irritability score (0.128), PANSS positive symptoms (0.115), heart rate variability (0.098), and medication adherence (0.087). Patients with three or more risk factors demonstrated 68.4% probability of aggression compared to 12.3% with zero factors. Physiological signals provided incremental predictive value beyond clinical and behavioural data alone (ΔAUC = 0.08, p