Abstract P63: Post-Stroke Cognitive Complaints and Normal MoCA Scores: A Possible Role for Machine Learning-Augmented Cognitive Screening

Abstract

Background: The Montreal Cognitive Assessment (MoCA) may be insufficiently sensitive to cognitive impairment in high-functioning stroke survivors. We examined the ability of a machine learning (ML) algorithm to distinguish young stroke survivors with mRS of 0-1 and MoCA >26 vs. age-matched controls. Methods: As part of a study comparing performance of the NIH Toolbox Cognitive Battery (NIHTB-CB) in characterizing cognitive deficits in young survivors, we assessed 52 survivors and 53 healthy controls. Voice recordings of subjects describing the Cookie Theft photo were analyzed using a Natural Language Processing algorithm incorporating 335 lexical and acoustic features. We used a stratified five-fold cross validation, performing a feature selection step before training where we selected for inclusion into the model the first k features with the highest absolute correlation with labels in the training fold. Area under the receiver operating curve (AUC) was calculated for each model. Results: MoCA was >26 in 83% of controls and 63% of stroke survivors. Using a Gaussian Naive Bayes Classifier, AUC for stroke survivors vs. controls in those with MoCA >26 was 0.74 (95%CI 0.58-0.91). Informative non-acoustic features included lexical and syntactic complexity, info-units (ie. features in the picture) and spatial orientation of info-units. The analysis is ongoing and relation with NIHTB-CB scores will be reported subsequently. Conclusions: There is a ceiling effect of the MoCA. Automated ML assessments may serve as efficient screening adjuncts prior to more detailed cognitive assessments in high-functioning stroke survivors with cognitive complaints. Further work is needed.