Development of Machine-Learning-Based Models for Detection of Cognitive Impairment in Patients Receiving Maintenance Hemodialysis

Background: Cognitive impairment is common but frequently undiagnosed in the dialysis population. We aimed to develop and validate a quick and accurate screening tool using machine-learning-based approaches in them. Methods: In this cross-sectional observational study, we administered the Mini-Mental State Examination (MMSE) and Cognitive Abilities Screening Instrument (CASI) in 508 hemodialysis patients and randomly divided them into a derivation set (70%) and a validation set (30%). Using three to five key items from MMSE and CASI as predictors, we developed six machine learning models, including Lasso, classification and regression tree (CART), random forest (RF), extreme gradient boosting, support vector machine (SVM), and artificial neural networks to identify those with a CASI score below the 20th percentile of age- and education-matched norms in the derivation set. We then evaluated the predictive performance of these models in the validation set. Results: The derivation samples (n = 357) had a mean (SD) age of 64.13 (11.92) years and a mean education level of 8.76 (4.91) years. Around 40% of participants had a CASI score below the 20th percentile. Among all models, the RF model achieved the highest performance of prediction, with an accuracy of 0.94, an area under the curve (AUC) of 0.95, and an F1 score of 0.92 in the validation set. The other models, except for CART, performed equally well in terms of AUC. Conclusions: Our study demonstrates that using machine-learning models, we can identify patients with impaired cognition with only several questions in CASI and MMSE within 5 min.