We conducted a retrospective cohort study utilizing electronic health record data and physiological monitoring from psychiatric inpatients admitted between January 2019 and December 2023. The study was approved by the institutional review board with waiver of informed consent for retrospective analysis of de-identified data. The study population comprised adults aged 18 - 85 years admitted to acute psychiatric units at three tertiary care hospitals. Site 1 was a 48-bed academic psychiatric inpatient unit in northern Italy (University Hospital setting, predominantly schizophrenia and bipolar disorder admissions). Site 2 was a 62-bed regional psychiatric hospital in central France (mixed voluntary and involuntary admissions, higher proportion of PTSD and borderline personality disorder). Site 3 was a 55-bed acute psychiatric unit in the United Kingdom (National Health Service, high comorbidity burden and substance use disorder prevalence). Incident reports were harmonized across sites using the Staff Observation Aggression Scale Revised (SOAS-R), which was adopted as the standardized incident documentation instrument at all three sites prior to the study period. Physiological monitoring was performed using the same wearable device model (Empatica E4 wristband) at all three sites, with a unified preprocessing pipeline applied centrally to raw sensor data before feature extraction. Nursing behavioural rating scales (MOAS, NOSIE-30) were translated and locally validated at each site; inter-site reliability was assessed on a 10% random sample of paired independent ratings, yielding ICC = 0.83 for irritability and ICC = 0.79 for social withdrawal, indicating acceptable cross-site reliability. Inclusion criteria required: 1) primary psychiatric diagnosis according to DSM-5 criteria; 2) admission to locked inpatient unit; 3) minimum length of stay 72 hours; and 4) availability of complete clinical and physiological data. Exclusion criteria included: 1) primary neurological disorder; 2) active delirium or severe cognitive impairment precluding assessment; 3) pregnancy; 4) medical conditions affecting autonomic function (severe diabetes, autonomic neuropathy); and 5) missing critical data elements.

We extracted data from four domains comprising 25 features. The index assessment time was defined as the nursing shift assessment at 08:00 on the day of each patient's third inpatient day, constituting the prediction timestamp. All features were extracted from data available strictly before this timestamp; the 7-day outcome window began at 08:01 on the same day. Look-back windows by domain were as follows. Clinical history features (prior aggression count, diagnosis, illness duration, lifetime hospitalizations, involuntary admissions, medication regimen) were extracted from the full electronic health record up to and including the day prior to the index assessment. Behavioural features (irritability score, sleep disturbance, medication adherence, substance use days, social withdrawal) were derived from nursing observations recorded in the 72-hour window preceding the index assessment, corresponding to the minimum length-of-stay inclusion criterion. Physiological signals (HRV RMSSD, skin conductance, morning cortisol, body temperature deviation) were extracted from the 24-hour wearable monitoring window immediately preceding the index assessment; morning cortisol was measured from a salivary sample collected at 07:00 on the index day. No feature was derived from data recorded after the index assessment timestamp. Incident reports used to construct the outcome variable were extracted from the 7-day window beginning at index assessment timestamp plus one minute.

Demographic characteristics included age, sex, and socioeconomic indicators [36][37].

Clinical history encompassed prior aggression episodes (lifetime count), primary diagnosis (schizophrenia, bipolar disorder, post-traumatic stress disorder, borderline personality disorder, major depressive disorder, substance use disorder), comorbidity count, illness duration, lifetime hospitalizations, involuntary admissions, and current medication regimen [38]-[40].

Behavioural patterns were assessed through nursing observations and patient reports including irritability score (0 - 10 scale), sleep disturbance (hours deviation from normal), medication adherence percentage, substance use days (past month), and social withdrawal (0 - 10 scale) [41][42].

Physiological signals were obtained through continuous wearable monitoring and laboratory assays: heart rate variability (root mean square of successive differences, RMSSD), skin conductance level (microSiemens), morning cortisol level (nmol/L), and body temperature deviation from baseline [43]-[45].

The primary outcome was occurrence of aggressive behaviours within 7 days of assessment, defined as any physical assault, weapon use, or severe verbal threat requiring intervention, documented through standardized incident reporting [46][47].

We evaluated four supervised learning algorithms: Random Forest, Gradient Boosting, Logistic Regression with L2 regularization, and Support Vector Machine with radial basis function kernel [48]-[50].

Random Forest was implemented with 200 estimators, maximum depth 10, minimum samples split 5, and balanced class weights to address outcome imbalance [51].

Gradient Boosting utilized 200 estimators, learning rate 0.1, maximum depth 5, and subsampling rate 0.8 [52].

Logistic Regression employed balanced class weights and maximum iterations 1000 [53].

Support Vector Machine used probability estimation with balanced class weights, C = 1.0, and automatic gamma scaling [54].

Feature preprocessing included median imputation for missing values (<5% of observations) and z-score standardization [55]. Data were partitioned into training (80%) and testing (20%) sets with stratification by outcome and diagnosis. Five-fold cross-validation was performed on the training set for hyperparameter optimization [56].

Primary performance metrics included area under the receiver operating characteristic curve (AUC-ROC), F1-score, sensitivity, specificity, and calibration assessed through Brier score [57]-[59]. Confidence intervals were calculated using 1000 bootstrap replications [60].

Feature importance was quantified using SHAP (SHapley Additive exPlanations) values, which provide consistent and locally accurate attribution of predictions to input features [61]. Partial dependence plots illustrated marginal effects of key features on predicted probability [62].

Heterogeneous treatment effects were analyzed across subgroups defined by diagnosis, age, sex, and baseline risk factors. Calibration was assessed through reliability diagrams comparing predicted probabilities to observed frequencies [63].

Baseline characteristics were compared between aggressive and non-aggressive groups using t-tests for continuous variables and chi-square tests for categorical variables. Model comparisons utilized DeLong’s test for correlated ROC curves [64]. Sensitivity analyses examined model performance under varying definitions of aggression (physical only vs. any aggression) and prediction windows (3-day vs. 7-day vs. 14-day) [65].

All analyses were performed using Python 3.9 with scikit-learn, XGBoost, and SHAP libraries. Two-tailed p-values < 0.05 were considered statistically significant.

Of 3847 patients screened, 2500 met inclusion criteria and were included in analysis (Figure 1Figure 1). Baseline characteristics are presented in Table 1. Mean age was 38.6 years (SD = 14.2), 55.0% were male, and 45.0% had schizophrenia as primary diagnosis. Prior aggression was reported in 62.0% of participants with mean 2.5 episodes (SD = 3.1). Aggressive behaviour occurred in 658 participants (26.3%) within the 7-day prediction window.

Statistical comparison between aggressive and non-aggressive patients revealed significant differences across multiple domains. Aggressive patients were younger (mean 36.8 ± 13.2 vs. 39.2 ± 14.5 years; t = 3.12, p = 0.002), more likely to be male (58.4% vs. 53.2%; χ² = 4.82, p = 0.028), and had higher rates of schizophrenia (53.2% vs. 42.1%; χ² = 18.42, p < 0.001). Historical aggression burden was substantially higher in the aggressive group (4.2 ± 3.8 vs. 1.8 ± 2.4 prior episodes; t = 15.24, p < 0.001).

Current clinical status differed markedly between groups. Irritability scores were more than twice as high in aggressive patients (6.8 ± 2.4 vs. 3.2 ± 2.1; t = 28.46, p < 0.001). Medication adherence was significantly lower (62.3 ± 22.1% vs. 78.4 ± 18.2%; t = 14.82, p < 0.001). Physiological markers showed autonomic dysregulation, with heart rate variability reduced by 26% (28.4 ± 9.8 vs. 38.2 ± 11.5 ms; t = 16.38, p < 0.001). Positive psychotic symptoms were more severe (PANSS positive: 24.8 ± 7.2 vs. 16.2 ± 5.8; t = 24.16, p < 0.001).

Figure 1. Psychiatric patient data distribution overview. (A) Age distribution by aggression risk status showing higher concentration of aggressive episodes in younger adults (peak 25 - 35 years); (B) Primary diagnosis distribution across the cohort (n = 2500), with schizophrenia (25.0%) and bipolar disorder (20.0%) as the most prevalent conditions; (C) Relationship between prior aggression episode count and current risk probability demonstrating exponential risk escalation beyond 3 episodes; (D - F) Box plots comparing physiological and behavioural measures by aggression status; (D) heart rate variability (median 26 vs. 37 ms); (E) irritability score (median 7 vs. 3), and (F) medication adherence (median 58% vs. 82%). All differences were significant at p < 0.001.

Table 1. Baseline characteristics by aggression status.

Data presented as mean ± standard deviation or n (%). HRV = heart rate variability; RMSSD = root mean square of successive differences; PANSS = Positive and Negative Syndrome Scale; CGI = Clinical Global Impression. p-values calculated using independent samples t-test for continuous variables and chi-square test for categorical variables.

The Random Forest classifier demonstrated superior predictive performance with AUC-ROC 0.84 (95% CI: 0.81 - 0.87), significantly exceeding other algorithms (Table 2, Figure 2Figure 2). This represents excellent discrimination according to standard interpretation (AUC > 0.80). Gradient Boosting achieved AUC 0.82 (95% CI: 0.79 - 0.85), also indicating excellent performance. Logistic Regression achieved moderate discrimination (AUC = 0.76, 95% CI: 0.72 - 0.79), while Support Vector Machine showed modest performance (AUC = 0.74, 95% CI: 0.71 - 0.78).

Cross-validation confirmed robustness with mean AUC 0.82 (SD = 0.04) for Random Forest, indicating minimal overfitting. The 95% confidence interval (0.81 - 0.87) excludes 0.50, confirming statistically significant discrimination (z = 12.84, p < 0.001). Calibration was acceptable with Brier score 0.142, indicating well-calibrated probability estimates (Figure 3Figure 3).

Table 2. Model performance comparison.

AUC = area under curve; CI = confidence interval; PPV = positive predictive value; NPV = negative predictive value. All metrics calculated on held-out test set (n = 500). Cross-validation AUC (5-fold): Random Forest 0.82 ± 0.04; Gradient Boosting 0.80 ± 0.03; Logistic Regression 0.75 ± 0.05; SVM 0.73 ± 0.06.

Diagnostic metrics for Random Forest at 0.5 threshold: true positives = 123, false positives = 39, true negatives = 311, false negatives = 27. Sensitivity = 82.0% (95% CI: 75.2 - 87.8%), specificity = 78.0% (95% CI: 73.4 - 82.2%), positive predictive value = 75.9%, negative predictive value = 83.6%, positive likelihood ratio = 3.73, negative likelihood ratio = 0.23.

Figure 2. ROC curves—Aggression risk prediction. Receiver operating characteristic curves comparing discriminative performance of four machine learning algorithms: Random Forest (AUC = 0.84; 95% CI: 0.81 - 0.87), Gradient Boosting (AUC = 0.82; 95% CI: 0.79 - 0.85), Logistic Regression (AUC = 0.76; 95% CI: 0.72 - 0.79), and Support Vector Machine with RBF kernel (AUC = 0.74; 95% CI: 0.71 - 0.78). The dashed diagonal line represents random classification (AUC = 0.500). Random Forest demonstrated significantly superior discrimination compared to all other models (DeLong’s test p < 0.001 vs. Logistic Regression and SVM; p = 0.08 vs. Gradient Boosting).

Figure 3. Calibration plots—Predicted vs. observed probabilities. Calibration plots for (A) Random Forest, (B) Gradient Boosting, (C) Logistic Regression, and (D) Support Vector Machine. Each plot shows the relationship between mean predicted probabilities (x-axis) and observed frequencies of aggression (y-axis) across 10 decile bins. The diagonal dashed line represents perfect calibration. Random Forest demonstrates excellent calibration (Brier score = 0.142, slope = 1.02, intercept = −0.01) with minimal deviation from ideal. Gradient Boosting shows good calibration (Brier = 0.156), while Logistic Regression and SVM modestly overestimate risk at higher probabilities.

SHAP (SHapley Additive exPlanations) analysis identified prior aggression count as the strongest predictor (mean |SHAP| = 0.142, 95% CI: 0.128 - 0.156), contributing 14.2% of total model prediction power. Current irritability score ranked second (mean |SHAP| = 0.128, 11.4% contribution), followed by PANSS positive symptoms (0.115, 10.2%), heart rate variability (0.098, 8.7%), and medication adherence (0.087, 7.8%).

Physiological features collectively contributed 24.3% of predictive power (HRV 8.7%, cortisol 5.2%, skin conductance 4.8%, temperature 3.2%, sleep disturbance 2.4%), demonstrating significant incremental value beyond clinical history and behavioural measures. Directionality analysis revealed that higher prior aggression (SHAP +0.142), irritability (+0.128), and positive symptoms (+0.115) increased predicted risk, while higher HRV (−0.098) and medication adherence (-0.087) were protective (Figure 4Figure 4).

Partial dependence plots revealed non-linear dose-response relationships (Figure 5Figure 5). Prior aggression showed exponential risk increase beyond 2 episodes (risk 18% at 0 episodes, 28% at 2 episodes, 52% at 5 episodes, 78% at 10 episodes). Irritability demonstrated threshold effects with minimal risk below score 4 (15%), accelerating increase between 4 - 7 (15% → 45%), and plateau above 7 (45% → 65%). Heart rate variability showed inverse linear relationship with risk doubling below 25 ms (45% vs. 22% at >35 ms).

Feature interaction analysis revealed significant multiplicative effects (Figure 6Figure 6). The combination of high prior aggression (≥5 episodes) with elevated irritability (≥7) produced supra-additive risk (predicted probability 78%) compared to main effects alone (aggression alone 45%, irritability alone 48%, expected additive 68%, interaction term + 10%). Medication adherence showed protective interaction with symptom severity, where excellent adherence (≥80%) reduced risk by

Figure 4. SHAP analysis of aggression risk prediction. (A) Feature importance ranking showing mean absolute SHAP values for top 15 features, coloured by direction of effect (red = increases risk, blue = decreases risk). Prior aggression count demonstrates the strongest influence (0.142), followed by irritability (0.128) and PANSS positive (0.115); (B) SHAP value distribution (beeswarm plot) for top 10 features, showing relationship between feature values (color: red = high, blue = low) and impact on prediction. High values of risk factors push predictions right (higher risk), while high values of protective factors push left; (C) Dependence plot for prior aggression count, coloured by irritability score tertile, revealing interaction where high irritability amplifies the effect of prior aggression; (D) Force plot showing feature contributions for a representative high-risk patient (baseline 0.263 → final 0.847), with prior aggression (+0.18), irritability (+0.15), and low HRV (+0.12) as primary drivers.

Figure 5. Partial dependence plots—Top 5 features. Marginal effects of the five most predictive features on aggression probability, holding other variables constant at their means: (A) Prior aggression count shows exponential increase beyond 2 episodes; (B) Irritability score demonstrates threshold effect with rapid escalation above score 6; (C) PANSS positive symptoms show linear increase with acceleration above score 25; (D) Heart rate variability shows inverse linear relationship with 2-fold risk increase below 25 ms; (E) Medication adherence demonstrates protective linear effect with 40% relative risk reduction from 20% to 100% adherence. Red dashed horizontal line indicates 50% probability threshold.

Figure 6. Feature interaction effects on aggression risk. (A) Interaction between prior aggression episode count (0, 2, 5, 10) and current irritability score. Lines show predicted risk across irritability range for each prior aggression level. Multiplicative interaction evident: patients with 10 prior episodes show steep risk escalation (25% → 85%) even at moderate irritability, while those with 0 episodes show minimal increase (10% → 25%); (B) Interaction between medication adherence (20%, 50%, 80%, 100%) and positive symptom severity (PANSS 7 - 49). Poor adherence combined with high symptoms creates multiplicative risk elevation (20% adherence + high PANSS = 68% risk), while excellent adherence demonstrates flattening protective effect (100% adherence reduces risk to 35% regardless of symptoms).

35% in high-symptom patients (PANSS > 25: 62% → 40%) but only 12% in low-symptom patients (PANSS ≤ 25: 18% → 16%).

Risk stratification using predicted probabilities demonstrated strong calibration across categories (Figure 7Figure 7). Very low risk (<20% probability) comprised 28.0% of the cohort (n = 700) with observed aggression rate 8.2% (57/700). Low risk (20% - 40%) included 35.0% (n = 875) with 18.4% rate (161/875). Moderate risk (40% - 60%) comprised 22.0% (n = 550) with 42.6% rate (234/550). High risk (60% - 80%) included 12.0% (n = 300) with 71.3% rate (214/300). Very high risk (>80%) comprised 3.0% (n = 75) with 89.5% rate (67/75).

The composite risk score combining five binary factors (prior aggression > 2, irritability > 6, PANSS positive > 25, substance use > 7 days, adherence < 50%) showed exponential cumulative effects. Zero factors: 12.3% risk (n = 380). One factor: 28.4% risk (n = 720). Two factors: 45.2% risk (n = 850). Three factors: 68.4% risk (n = 400). Four or five factors: 82.1% risk (n = 150).

Figure 7. Risk stratification and threshold optimization. (A) Distribution of predicted probabilities by actual aggression status, showing clear separation with minimal overlap (Kolmogorov-Smirnov D = 0.62, p < 0.001). Vertical dashed line indicates 0.5 classification threshold; (B) Actual aggression rates across five risk categories derived from predicted probability deciles, demonstrating strong calibration (Hosmer-Lemeshow χ² = 6.82, p = 0.56); (C) Cumulative risk factor analysis showing exponential increase in aggression rate with number of present risk factors (odds ratio per additional factor = 2.84, 95% CI: 2.46 - 3.28); (D) Threshold analysis curves: sensitivity decreases (0.95 → 0.45), specificity increases (0.45→0.95), and F1-score peaks at 0.5 threshold (0.78), supporting this as optimal operating point.

Threshold optimization analysis identified 0.50 as optimal for balanced sensitivity-specificity. At this threshold: sensitivity 82.0%, specificity 78.0%, F1-score 0.78. Lowering threshold to 0.30 increased sensitivity to 92% but reduced specificity to 62% (higher false positives). Raising threshold to 0.70 increased specificity to 91% but reduced sensitivity to 64% (higher false negatives) (See Figure 8Figure 8).

Figure 8. Clinical decision support visualizations. (A) Risk by primary diagnosis shows the highest rates in mixed personality disorder (35%) and borderline personality disorder (32%), the lowest in major depression (22%) and substance use disorder (24%). Color coding: green (<25%), orange (25% - 30%), red (>30%); (B) Medication adherence demonstrating approximately linear protective effect: each 20% improvement in adherence reduces risk by ~8% (from 45% at 0% adherence to 15% at 100%); (C) Heart rate variability risk stratification: very low HRV (<20 ms) associated with 3-fold increased risk (42%) compared to normal HRV (>50 ms, 15%). Blue dashed line indicates population mean (26.3%); (D) Patient distribution by cumulative risk factor count: 15% have 0 factors, 29% have 1 factor, 34% have 2 factors, 16% have 3 factors, 6% have 4 - 5 factors.

Model performance was consistent across diagnostic groups with AUC ranging from 0.81 (substance use disorder) to 0.87 (schizophrenia) (Figure 9Figure 9). No significant heterogeneity was observed (I² = 12%, p = 0.34). Performance was maintained in patients with comorbid personality disorders (AUC = 0.83) and substance use (AUC = 0.81). Age-stratified analysis showed optimal discrimination in young adults 18 - 30 years (AUC = 0.86), with stable performance across middle age 31 - 60 years (AUC = 0.84 - 0.85) and older adults 61 - 75 years (AUC = 0.82). Sex-specific models showed comparable performance: male (AUC = 0.85), female (AUC = 0.84), other gender identities (AUC = 0.83).

Sensitivity analyses confirmed robustness. Restricting outcome to physical

Figure 9. Model performance across patient subgroups. AUC-ROC scores for the Random Forest model stratified by (A) primary diagnosis, (B) age group, and (C) sex. Performance consistently exceeded 0.80 across all subgroups. Highest discrimination observed in schizophrenia (0.87) and bipolar disorder (0.85); lowest in substance use disorder (0.81) and major depression (0.82). Age analysis shows peak performance in young adults (18 - 30: 0.86) with gradual decline in older age groups (76+: 0.81). Sex differences minimal (male 0.85, female 0.84, other 0.83). Red dashed lines indicate AUC = 0.80 threshold for good discrimination.

aggression only (excluding verbal threats, n = 412 events) yielded similar performance (AUC = 0.82, 95% CI: 0.79 - 0.85). Extending prediction window to 14 days reduced discrimination (AUC = 0.78, 95% CI: 0.75 - 0.81) due to temporal dilution, while shortening to 3 days improved AUC to 0.87 (95% CI: 0.84 - 0.90) but reduced clinical utility. Excluding physiological variables (HRV, cortisol, skin conductance) reduced AUC by 0.08 (ΔAUC = 0.08, 95% CI: 0.05 - 0.11, p < 0.001), confirming significant incremental value of biomarkers. Using only clinical history variables (excluding behavioural and physiological) reduced AUC by 0.15 (AUC = 0.69), demonstrating the importance of multi-modal integration.

This study demonstrates that a multi-modal machine learning framework achieves substantially improved prediction of short-term aggression risk in psychiatric inpatients compared to traditional assessment methods. The Random Forest classifier attained AUC-ROC 0.84, representing excellent discrimination that exceeds performance of established instruments such as the HCR-20 (typically AUC 0.65 - 0.75) and Broset Violence Checklist (AUC 0.60 - 0.70) [66]-[68]. The integration of physiological biomarkers provided significant incremental predictive value beyond clinical and behavioural data alone.

The identification of prior aggression, current irritability, and positive psychotic symptoms as strongest predictors aligns with established clinical knowledge while providing quantitative precision [69][70]. The novel contribution of heart rate variability as a top five predictor supports theoretical models linking autonomic dysregulation to behavioural dyscontrol [71][72]. These findings suggest that objective physiological monitoring can capture preclinical states of arousal dysregulation that precede observable behavioural disturbances.

The risk stratification framework enables tiered intervention strategies. Patients in very high-risk category (>80% probability) may warrant intensive monitoring, prophylactic medication adjustment, and environmental modifications [73][74]. Moderate-risk patients may benefit from enhanced observation and behavioural interventions. Low-risk classification could support decisions to reduce restrictive measures, potentially improving therapeutic alliance and autonomy [75][76].

The explicit quantification of prediction uncertainty through calibrated probability estimates supports shared decision-making between clinicians, patients, and families [77][78]. Unlike binary risk categorization, probability-based communication facilitates nuanced discussions about precautionary measures while avoiding deterministic labelling that may exacerbate stigma [79][80].

Prior machine learning studies in psychiatric aggression prediction have focused primarily on static demographic and historical factors, achieving AUC values of 0.65 - 0.75 [81]-[83]. Our incorporation of dynamic behavioural and physiological variables represents methodological advancement. Recent studies utilizing electronic health record data have shown similar performance to our clinical history model but lacked physiological integration [84][85].

The emphasis on short-term (7-day) prediction distinguishes this work from studies predicting violence over months or years [86][87]. Short-term prediction aligns with clinical workflows where immediate risk management decisions are required, though it necessitates frequent model updating as patient states evolve [88][89].

Several limitations warrant consideration. The retrospective design precludes assessment of real-time implementation challenges including data latency, sensor reliability, and alert fatigue [90][91]. Prospective validation in operational settings is essential before clinical deployment. The relatively high prevalence of aggression (26%) in our cohort may reflect selection bias toward higher-acuity units, potentially affecting calibration in lower-prevalence settings [92][93].

Physiological monitoring requires patient cooperation and functional sensors, which may be challenging in severely agitated or uncooperative patients precisely when risk assessment is most critical [94][95]. Missing data imputation strategies, while conservative, may not fully capture information loss in incomplete observations [96][97].

The generalizability across healthcare systems with varying resources, cultural contexts, and patient populations requires evaluation [98][99]. Model performance in community hospitals, forensic settings, and adolescent or geriatric populations remains to be established [100][101].

Integration with electronic health record systems and real-time alert infrastructure represents the immediate next step [102][103]. Development of closed-loop systems that automatically adjust monitoring intensity based on risk level could optimize resource allocation [104][105]. Prospective randomized trials comparing ML-guided versus standard risk management are needed to demonstrate impact on patient outcomes and safety events [106][107].

Expansion to prediction of other adverse events including self-harm, elopement, and suicide attempts would enhance clinical utility [108][109]. Incorporation of natural language processing of clinical notes and patient communications may capture additional predictive signals [110][111]. Federated learning approaches enabling multi-site model development without data sharing could address privacy concerns and improve generalizability [112][113].

Interpretability enhancements including natural language generation of risk explanations and visualization tools for frontline staff may facilitate adoption [114][115]. Ethical frameworks governing appropriate use of predictive algorithms in psychiatric settings, including transparency, accountability, and protection from discriminatory application, require development alongside technical advances [116][117].