The performance of machine learning models in predicting postpartum depression: a meta-analysis and systematic review

To evaluate the effectiveness of machine learning (ML) approaches in predicting individuals with postpartum depression (PPD), this study systematically reviewed and meta-analysed existing evidence.

A systematic search was conducted across five databases including Cochrane Library, PsycINFO, Embase, Medline and the China National Knowledge Infrastructure. The performance metrics of ML models were pooled and risk of publication bias was assessed. The quality of the studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool.

From an initial 4,994 identified articles, 10 studies involving a total of 9,189 participants met the inclusion criteria. The findings revealed high overall predictive performance: the pooled Area Under the Curve (AUC) was 0.889, pooled accuracy was 0.850 (95% CI: 0.800-0.899), pooled sensitivity was 0.706 (95% CI: 0.589-0.801), and pooled specificity was 0.886 (95% CI: 0.833-0.921). Significant heterogeneity was observed, with an I² value of 98.24% and a Q statistic of 513.03 (p < 0.001). The Deek's funnel-plot asymmetry test showed that there was no significant publication bias (p = 0.919). Subgroup analysis indicated higher sensitivity when diagnostic tools were used compared to screening tools.

This meta-analysis synthesised evidence on the prediction of PPD using ML approaches. Although ML models for PPD prediction exhibit high specificity, their limited sensitivity hinders overall predictive accuracy.