Machine learning algorithms for predicting PTSD: a systematic review and meta-analysis.

Area of Science:

• Computational psychiatry
• Psychological assessment
• Machine learning in healthcare

Background:

• Post-traumatic stress disorder (PTSD) poses a significant public health challenge.
• Accurate predictive models are crucial for early intervention and risk stratification.
• Existing machine learning (ML) models for PTSD prediction vary in performance and reliability.

Purpose of the Study:

• To systematically review and meta-analyze the predictive accuracy of ML models for PTSD.
• To evaluate the risk of bias (ROB) in studies developing and validating these PTSD predictive models.
• To identify the most effective ML algorithms and data types for PTSD prediction.

Main Methods:

• Systematic review and random-effect meta-analysis of 23 studies.
• Inclusion of diverse samples predicting PTSD using ML algorithms.
• Performance pooled using area under the curve (AUC); ROB appraised using the PROBAST tool.

Main Results:

• 48% of included studies had high ROB, with remaining having unclear ROB.
• Tree-based models demonstrated promising pooled AUCs across various trauma types (e.g., sexual trauma: 0.861, firefighters: 0.96).
• Significant study variability and lack of external validation limited applicability.

Conclusions:

• ML models, particularly tree-based algorithms, show potential for PTSD prediction.
• High ROB and insufficient external validation necessitate cautious interpretation of current findings.
• Future research should prioritize adherence to reporting standards and robust external validation for clinical utility.