Machine learning models show moderate predictive accuracy for CBT depression outcomes in external validation

Photo by Steve A Johnson / Unsplash

Behaviour research and therapy Published June 1, 2026 Medically reviewed June 1, 2026 Study authors: Lees Jennifer, Delgadillo Jaime PubMed ↗ DOI ↗ By Dr. Ji-eun Park, MD · Brain, Mind & Pain

Key Takeaway

Consider that machine learning models show moderate predictive accuracy for CBT depression outcomes, but technical reporting is sparse.

This is a systematic review and meta-analysis examining the use of machine learning (ML) methods to predict depression treatment outcomes in patients receiving cognitive behavioural therapy (CBT). The synthesis included 11,733 patients across studies.

The key finding is that the generalisability of predictions from ML models to external validation samples was replicated in 5 out of 7 studies. The pooled effect size was a correlation of r = 0.45, with a 95% CI of 0.26 to 0.63 and a p-value of less than .001, indicating a positive association.

A major limitation noted by the authors is that reporting of technical details of ML model-training methods is generally sparse. The review does not report on safety, adverse events, or clinical efficacy beyond prediction.

Practice relevance was not reported. The findings suggest ML may have a role in predicting outcomes, but the evidence is limited by methodological reporting gaps and should not be overstated as demonstrating clinical efficacy or safety.

Study Details

Study typeMeta analysis

EvidenceLevel 1

PublishedJun 2026

PMID41930537

View Original Abstract ↓

BACKGROUND: Cognitive behavioural therapy (CBT) is an empirically-supported treatment for depression, although some patients respond well and others do not. The use of machine learning (ML) could potentially help to predict which patients may benefit most from CBT. OBJECTIVES: To synthesise the results of CBT studies applying ML to predict depression treatment outcomes. METHODS: Systematic searches were conducted across three databases and eligible studies were assessed for risk of bias. ML methods and predictor variables were summarised using a narrative synthesis. Predictive performance indices were harmonized into a common effect size (r) and pooled using random-effects meta-analysis, separately for evaluations using [1] internal and [2] external cross-validation. RESULTS: Twenty-four studies (n = 11,733) met eligibility criteria, of which only four (16.7%) were deemed at low risk of bias. The most commonly selected predictors were depression-related features (e.g., severity, symptom subtypes), functional/social impairment, sociodemographic characteristics (e.g., employment), comorbidity (mental/personality disorders) and adversity (current and past events). Studies with the most rigorous cross-validation methodology show replicated evidence that predictions from ML models generalise to external validation samples (in 5 out of 7), with a moderate effect size (r = 0.45; 95% CI: 0.26 to 0.63; p < .001). LIMITATIONS: Reporting of technical details of ML model-training methods is generally sparse. CONCLUSION: Replicated evidence indicates that ML methods can predict depression treatment outcomes, with adequate generalisability to new samples.