Clinical prediction models for post-stroke depression: a systematic review and meta-analysis

Abstract

Background: Post-stroke depression (PSD) is a prevalent neuropsychological consequence of stroke, associated with cognitive decline, disability, and increased mortality. Early prediction of PSD is critical for timely interventions and better outcomes. This study evaluates the effectiveness of various clinical prediction models, particularly machine learning methods, in forecasting PSD.

Methods: A systematic review and meta-analysis were conducted to evaluate predictive models for post-stroke depression (PSD) using data from 16 studies. The databases searched included PubMed, Embase, Cochrane Library, and Web of Science, covering publications from the year 2000 to the present. The risk of bias in the predictive models was assessed using the Prediction model Risk of Bias Assessment Tool+AI (PROBAST+AI). The review encompassed both traditional statistical methods and machine learning algorithms. Predictive accuracy was analyzed through the area under the curve (AUC) values of these models, considering various data sources, such as clinical, cognitive, and biomarker data. R packages, including 'metafor,' 'meta,' and 'forestplot,' were used for the analysis.

Results: 16 studies were included. Neural Network models yielded the highest pooled AUC (0.88, 95% CI: 0.45-0.98), although this estimate was based on only two studies and exhibited wide confidence intervals. Logistic Regression, Decision Tree, and K-Nearest Neighbor models showed comparable predictive ability, with pooled AUC values ranging from 0.77 to 0.83. Support Vector Machine models demonstrated the lowest predictive performance (AUC = 0.68) but exhibited the lowest heterogeneity. Among different data sources, functional, physical, and cognitive assessments yielded the highest predictive accuracy (AUC = 0.86, 95% CI: [0.81, 0.90]), followed by biomarker-based models (AUC = 0.80, 95% CI: [0.71, 0.86]). Within retrospective studies, biomarker-based data sources demonstrated significantly superior predictive performance (AUC = 0.94, 95% CI: 0.92-0.96).

Conclusions: Machine learning models, particularly Neural Networks, show potential for predicting post-stroke depression, although current evidence is limited by small sample sizes and high heterogeneity. Traditional approaches such as Logistic Regression and Decision Tree models also demonstrate stable and competitive performance. Among data sources, functional, physical, and cognitive assessments provide the strongest predictive value, while biomarker-based models appear particularly effective in retrospective analyses. Despite these findings, the limited number of high-quality studies and methodological inconsistencies highlight the need for rigorous, prospective, and multicenter validation to establish reliable and generalizable predictive models for post-stroke depression.

Systematic review registration: https://www.crd.york.ac.uk/prospero/, identifier CRD42025635227.

Keywords: PROBAST+AI; artificial intelligence; depression prediction; machine learning models; meta-analysis; post-stroke depression; systematic review.