Detection of Cognitive Impairment From eSAGE Metadata Using Machine Learning

Abstract

Objective: Using the metadata collected in the digital version of the Self-Administered Gerocognitive Examination (eSAGE), we aim to improve the prediction of mild cognitive impairment (MCI) and dementia (DM) by applying machine learning methods. Patients and Methods: A total of 66 patients had a diagnosis of normal cognition (NC), MCI, or DM, and eSAGE scores and metadata were used. eSAGE scores and metadata were obtained. Each eSAGE question was scored and behavioral features (metadata) such as the time spent on each test page, drawing speed, and average stroke length were extracted for each patient. Logistic regression (LR) and gradient boosting models were trained using these features to detect cognitive impairment (CI). Performance was evaluated using 10-fold cross-validation, with accuracy, precision, recall, F1 score, and receiver operating characteristic area under the curve (AUC) score as evaluation metrics. Results: LR with feature selection achieved an AUC of 89.51%, a recall of 87.56%, and an F1 of 85.07% using both behavioral and scoring. LR using scores and metadata also achieved an AUC of 84.00% in detecting MCI from NC, and an AUC of 98.12% in detecting DM from NC. Average stroke length was particularly useful for prediction and when combined with 4 other scoring features, LR achieved an even better AUC of 92.06% in detecting CI. The study shows that eSAGE scores and metadata are predictive of CI. Conclusions: eSAGE scores and metadata are predictive of CI. With machine learning methods, the metadata could be combined with scores to enable more accurate detection of CI.