Machine learning-based prediction of 3-6-month post-stroke cognitive impairment using acute-phase clinical data: a...

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Post-stroke cognitive impairment (PSCI) is a frequent and debilitating consequence of stroke.
• Early identification of patients at high risk for PSCI remains a significant clinical challenge.
• Developing accurate prognostic models is crucial for timely intervention and management.

Purpose of the Study:

• To develop and externally validate a machine learning model for predicting PSCI risk.
• To utilize acute-phase predictors (within 1 week post-stroke) for risk estimation.
• To assess the model's performance and interpretability for clinical application.

Main Methods:

• Retrospective analysis of two inpatient rehabilitation cohorts (n=1,369 and n=341).
• Feature selection using Boruta and LASSO, with eight algorithms trained via 10-fold cross-validation.
• Performance evaluation using AUROC, average precision, accuracy, calibration metrics, and decision-curve analysis (DCA).

Main Results:

• The XGBoost model achieved an AUROC of 0.827 in external validation, with good accuracy (0.8328) and specificity (0.9241).
• Key predictors included diabetes, NIHSS, Hgb, Hcy, apoA1, TG, ALB, WMH, smoking, and age.
• The model demonstrated acceptable calibration and positive net benefit in DCA within a specific probability range.

Conclusions:

• The developed machine learning model shows potential for PSCI risk stratification in post-stroke rehabilitation patients.
• The model's interpretability through SHAP values aids in understanding predictive factors.
• Further validation in diverse stroke populations is necessary for broader clinical implementation.