Using Machine Learning to Predict Cognitive Impairment From Functional Assessments: A Study of 274 Stroke Patients

Abstract

Cognitive impairment, which commonly occurs after stroke, adversely impacts rehabilitation outcomes. The administration of traditional screening tools such as the Mini-Mental State Examination (MMSE) can be challenging for patients with communication or motor limitations. This study aimed to determine whether routinely collected functional assessments can effectively predict cognitive impairment in stroke patients undergoing robotic-assisted gait training with the Lokomat system. In this retrospective analysis, data from 274 stroke patients who received Lokomat-based rehabilitation were examined. Patients with complete data on the MMSE, Fugl-Meyer Assessment (FMA), Modified Barthel Index (MBI), and Timed Up and Go (TUG) test were included in the final analysis. Cognitive impairment was defined as an MMSE score < 24. Five supervised machine learning models—logistic regression, decision tree, random forest, support vector machine (SVM), and XGBoost—were developed to classify cognitive status using FMA, MBI, and TUG as predictors. The performance of each model was assessed by the area under the curve (AUC), accuracy, sensitivity, and precision. Logistic regression demonstrated the highest predictive performance (AUC = 0.92, accuracy = 89.5%), followed by random forest (AUC = 0.88) and XGBoost (AUC = 0.87). FMA was identified as the most significant predictor in tree-based models. Patients classified as cognitively impaired exhibited significantly lower functional scores than their unimpaired counterparts (p < 0.001). Functional assessments routinely performed during stroke rehabilitation can provide a practical and accurate method for identifying cognitive impairment, especially when cognitive screening is constrained. These results support the incorporation of machine learning–based tools into clinical decision-making during robotic rehabilitation.