Development and validation of machine learning models to predict PTSD at multiple time points in hospitalized trauma patients

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: AUC 0.84 (1 month, SVM), 0.78 (3 months, Random Forest), 0.87 (6 months, Random Forest)

• Primary Outcome: PTSD development at 1, 3, and 6 months post-trauma

• Nnt Or Sensitivity: AUC values indicate good to excellent discrimination (>0.75)

• Confidence Interval: Not reported in abstract

Clinical Application

High feasibility as models use routine clinical biochemical markers already collected in standard care Highly relevant to Australian trauma centers and emergency departments. Could integrate with existing trauma registries and support early mental health screening initiatives aligned with RACGP mental health guidelines Hospitalized trauma patients in acute care settings, particularly relevant for emergency departments and trauma centers

Abstract

BACKGROUND: Posttraumatic stress disorder (PTSD) is a severe trauma-related mental disorder with high global burden. Early identification remains challenging, particularly in developing regions. Leveraging routinely collected clinical biomarkers may offer a scalable solution for timely risk detection. METHOD: This multicenter study utilized two independent prospective trauma cohorts to develop and externally validate machine learning models for predicting PTSD at 1, 3, and 6 months post-trauma. The development cohort consecutively enrolled 382 hospitalized trauma patients (Nov 2023-Dec 2024), while the external validation cohort included 204 patients (Oct 2021-Nov 2022). Predictors comprised psychosocial variables and routine clinical biochemical markers, including inflammatory indicators, liver and renal function markers, and hematological parameters. Five models were constructed (Logistic Regression, Random Forest, SVM, XGBoost, and Neural Network), with performance evaluated by AUC and feature importance interpreted via SHAP. RESULTS: Support Vector Machine achieved the best performance at 1 month (AUC = 0.84), while Random Forest models performed best at 3 and 6 months (AUC = 0.78 and 0.87). SHAP analysis indicated that ASDS were the most important predictors at all time points, with feature importance shifting over time. Acute stress and psychological state were dominant at 1 month, fearful experiences with physiological responses at 3 months, and chronic psychological burden and emotional state at 6 months. CONCLUSIONS: PTSD can be reliably predicted using psychosocial data and routine clinical biochemical markers. These findings support the integration of clinically accessible biological data into early PTSD risk screening, particularly in resource-limited settings.

References

1. 1.Chu, X., Dai, X., Jiang, X., Liu, S., Chen, W., Zhao, S., Zhou, Y., & Shi, X. (2026). Development and validation of machine learning models to predict PTSD at multiple time points in hospitalized trauma patients. Journal of Affective Disorders, 121096. https://doi.org/10.1016/j.jad.2025.121096