Document resource
Background Machine learning (ML) has the potential to improve prediction of treatment outcomes and advance precision mental health care. However, relatively little research has applied ML methods to predict the outcome of post-traumatic stress disorder (PTSD), and previous systematic reviews of ML psychological therapy research have found that methodological and reporting limitations are common. Therefore, the aim of this study was to systematically review studies that applied ML methods to predict outcomes of psychological therapies for PTSD and evaluate their methodological rigour.Methods This systematic review was pre-registered with PROSPERO (CRD42022325021) and followed PRISMA guidelines. Inclusion criteria were any studies that applied ML methods to predict outcomes (e.g., symptom change, dropout) of psychological therapies recommended by clinical practice guidelines for the treatment of current PTSD in adults (e.g., trauma-focussed CBT, prolonged exposure, EMDR). Four databases were searched (APA PsycInfo, PTSDpubs, PubMed, Scopus) using pre-registered search terms, forward and backward citation searches performed, and eligible study authors contacted to request potentially eligible studies. Risk of bias was assessed using PROBAST. Study methods and findings were narratively synthesized, and adherence to ML best practices evaluated. Heterogeneity of study methods precluded meta-analysis of model prediction accuracy.Results 1,570 unique titles and abstracts were screened, 48 reports assessed for eligibility, and 17 studies met the pre-registered inclusion criteria. Studies applied a diverse range of ML methods for different purposes. Fourteen studies used supervised ML methods, eight of which used decision tree-based methods, six of which used ensemble tree methods such as random forest and boosting algorithms. Five studies used unsupervised ML methods, three of which used k-means clustering. Studies performed retrospective and prospective analyses of data from randomised trials, cohort studies, and routine practice. Most studies used clinical, demographic, and/or psychometric data; four studies used neuroimaging data. The number of participants with the outcome in a training sample ranged from < 36 to 397. All studies were high risk of bias. None had an appropriate sample size or performed sample size calculation, and eight studies did not report model evaluation metrics. Regarding ML best practices, seven studies did not report hyperparameter setting, six did not report internal cross-validation, and only one performed external validation. Four studies compared ML and traditional statistical methods against one another, with mixed results.Conclusion ML has the potential to advance precision treatment for PTSD, but to properly explore this potential ML methods must be applied with greater adherence to best practice guidelines.