Can Machine Learning predict therapeutic outcomes in affective and not affective psychosis? A systematic review and meta-analysis

Machine learning (ML) could be useful in identifying reliable predictors of treatment response in affective and not affective psychoses, potentially helping to propose personalized interventions. In this systematic review and meta-analysis, we evaluated studies exploiting ML algorithms to predict the improvement of psychotic symptoms, cognition and quality of life in psychoses related to different treatments. We searched MEDLINE (PubMed), Web of Science, and PsycINFO databases updated until February 2024, identifying 64 articles published in English in peer-reviewed journals. We modelled a random-effects meta-analysis to estimate the overall accuracy reached in 51 studies. Subgroup analyses and meta regressions were performed to compare predictive accuracy across different predicted target class (i.e., improvers or responders versus not responders or treatment-resistant), diagnosis, input features, type and duration of treatments, ML algorithms, sample size, year of publication and quality assessment, evaluated with the PROBAST tool. ML models predicted a treatment response with a total accuracy of 80 % (95 %CI [0.76;0.83]), despite detecting a high heterogeneity (I²=0.89). Significant differences were observed between input features (p = .004) and treatments (p = .01). The best predictor was electroencephalography data (88 % of accuracy, 95 %CI [0.82;0.93], I²=0.50), followed by the combined treatments (85 % of accuracy, 95 %CI [0.82;0.87], I²=0.51). We identified a general low quality of studies, with 44 having a high risk of bias. Overall, ML seems a promising tool for predicting therapeutic outcomes in affective and not affective psychoses. However, specific attention should be paid to enhancing reproducibility and improving study methodology to better translate results into clinical practice.