Prediction of post stroke depression with machine learning: A national multicenter cohort study

Area of Science:

• Neurology
• Psychiatry
• Data Science

Background:

• Post-stroke depression (PSD) is a frequent complication with low detection rates.
• Existing prediction models often lack comprehensive data, limiting clinical utility.

Purpose of the Study:

• To integrate multimodal data (clinical, biomarker, neuroimaging) for PSD prediction.
• To validate machine learning models for identifying high-risk PSD patients.

Main Methods:

• A multicenter cohort of 4298 acute ischemic stroke (AIS) patients was analyzed.
• Four machine learning models were implemented and compared.
• A Gradient Boosting Decision Tree (GBDT) model was developed using clinical, biomarker, and neuroimaging data.

Main Results:

• The GBDT model achieved high predictive performance (AUC 0.8626 in test set, 0.8185 in external validation).
• Key predictors included NIHSS score, lesion location (left-sided), lacunar infarcts, homocysteine, and systolic blood pressure.
• The GBDT model outperformed other tested machine learning algorithms.

Conclusions:

• Machine learning models show strong potential for predicting PSD.
• Identifying high-risk patients (e.g., those with high NIHSS, specific lesion types, elevated HCY, high SBP) allows for personalized management.
• Early and precise interventions can help prevent or delay PSD onset.