Accuracy of Machine Learning in Predicting Post-Stroke Depression: A Systematic Review and Meta-Analysis.

INTRODUCTION

Post-stroke depression is one of the important complications of stroke and affects patients' quality of life. Early identification of post-stroke depression is crucial for its timely prevention. The accuracy of machine learning as a prediction method is controversial. To systematically analyze these studies, we conducted a systematic evaluation to review the effectiveness of the machine learning prediction models in predicting post-stroke depression based on meta-analysis.

METHODS

As of November 20, 2023, we conducted a systematic literature retrieval in databases such as PubMed, Embase, Cochrane, and Web of Science to investigate machine learning model predictions in patients who had post-stroke depression. The tool used in this study to assess the quality of the retrieved literature is PROBAST (Prediction model Risk of Bias Assessment Tool).

RESULTS

A total of 28 studies with 85,223 patients as subjects were included in the meta-analysis. According to data, the c-index in the validation set was significantly lower than that in the training set and may have been at risk of overfitting, even though there was a desirable accuracy. In addition, we found substantial differences in the duration of follow-up for post-stroke depression, and meta-regression showed that the c-index based on the prediction models did not decay with longer follow-up.

CONCLUSIONS

Reasonable prediction models are effective prediction tools for post-stroke depression. A reasonable prediction seems to predict the risk of post-stroke depression occurrence at different time points and can provide a prevention tool specific to the risk.