Development and validation of a machine learning-based risk prediction model for post-stroke cognitive impairment.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Post-stroke cognitive impairment (PSCI) affects a significant portion of stroke survivors.
• Current machine learning (ML) models for predicting PSCI require optimization for clinical utility.
• Early and accurate prediction of PSCI is crucial for timely intervention and improved patient prognosis.

Purpose of the Study:

• To develop and validate a reliable ML-based predictive model for PSCI in a Chinese population.
• To identify significant clinical predictors associated with PSCI development post-acute ischemic stroke (AIS).
• To compare the performance of various ML algorithms in predicting PSCI.

Main Methods:

• Data from 494 AIS patients were collected, with cognitive function assessed using MMSE or MOCA scores.
• The Least Absolute Shrinkage and Selection Operator (LASSO) and logistic regression (LR) were used for feature selection from 49 clinical parameters.
• Seven ML models, including XGBoost, were trained and validated using tenfold cross-validation, with performance metrics such as AUROC, accuracy, and F1 score evaluated.

Main Results:

• PSCI was present in 58.50% of the studied AIS patients.
• Key predictors identified for PSCI include age, NIHSS, HAMD-24, PSQI, ALB, FBG, hypertension, paraventricular lesion, and number of lesions.
• The XGBoost model achieved the highest AUROC of 0.980, outperforming other evaluated ML models.

Conclusions:

• The developed XGBoost model demonstrates high accuracy in predicting PSCI 3-6 months post-stroke.
• The model effectively utilizes identified clinical predictors to identify patients at risk of cognitive impairment.
• This ML tool facilitates early identification of at-risk individuals, paving the way for timely clinical interventions and management strategies.