Quantitative forecasting of PTSD from early trauma responses: a Machine Learning application.

Area of Science:

• Computational psychiatry
• Clinical informatics
• Psychotraumatology

Background:

• Predicting mental disorder trajectories from early data is challenging due to computational limitations.
• Early identification of trauma survivors at risk for post-traumatic stress disorder (PTSD) is feasible due to salient onset and available risk indicators.

Purpose of the Study:

• Evaluate Machine Learning (ML) forecasting for identifying and integrating unique predictive characteristics.
• Determine the accuracy of ML in forecasting non-remitting PTSD within 10 days of a traumatic event.

Main Methods:

• Collected data from 957 trauma survivors on event characteristics, emergency department observations, and early symptoms over 15 months.
• Employed ML feature selection to identify key predictors and Support Vector Machines (SVMs) for classification.
• Compared prediction accuracy using ML-selected features, all features, and Acute Stress Disorder (ASD) symptoms alone.

Main Results:

• An ML feature selection algorithm identified 16 key predictors.
• Predicting non-remitting PTSD using selected features (AUC = .77) was comparable to using all available information (AUC = .78).
• Predicting PTSD status (AUC = .71) was less accurate than predicting membership in a non-remitting trajectory (AUC = .71), while ASD symptoms alone performed poorly (AUC = .60).

Conclusions:

• ML methods show promise in forecasting PTSD by identifying and integrating unique risk indicators.
• This approach can develop algorithms for probabilistic risk assessment of chronic posttraumatic stress psychopathology.
• ML facilitates the integration of complex biological, psychological, and social information for improved PTSD prediction.