Using machine learning methods to predict the outcome of psychological therapies for post-traumatic stress disorder: A systematic review

Background

A number of treatments are available for post-traumatic stress disorder (PTSD), however, there is currently a lack of data-driven treatment selection and adaptation methods for this condition. Machine learning (ML) could potentially help to improve the prediction of treatment outcomes and enable precision mental healthcare in practice.

Objectives

To systematically review studies that applied ML methods to predict outcomes of psychological therapy for PTSD in adults (e.g., change in symptoms, dropout rate), and evaluate their methodological rigour.

Methods

This was a pre-registered systematic review (CRD42022325021), which synthesised eligible clinical prediction studies found across four research databases. Risk of bias was assessed using the PROBAST tool. Study methods and findings were narratively synthesised, and adherence to ML best practice evaluated.

Results

Seventeen studies met the inclusion criteria, including samples derived from experimental and observational study designs. All studies were assessed as having a high risk of bias, notably due to inadequately powered samples and a lack of sample size calculations. Training sample size ranged from N < 36–397. The studies applied a diverse range of ML methods such as decision trees, ensembling and boosting techniques. Five studies used unsupervised ML methods, while others used supervised ML. There was an inconsistency in the reporting of hyperparameter tuning and cross-validation methods. Only one study performed external validation.

Conclusions

ML has the potential to advance precision psychotherapy for PTSD, but to enable this, ML methods must be applied with greater adherence to best practice guidelines.