P3–219: A machine-learning approach for integration of computerized cognitive data in the neuropsychological assessment of older adults

Background: In progressive conditions like dementia, it is important to detect subtle cognitive changes early and initiate intervention to slow or even halt disease progression. Essential to early detection is comprehensive cognitive assessment. Computerized neuropsychological batteries offer rapid, standardized, and precise assessment, as well as automatic scoring, but it remains a formidable task to optimally integrate the information. This study aims to identify a machine learning (ML) algorithm that optimally and automatically integrates all available computerized cognitive testing scores to improve neuropsychological assessment of older adults. Methods: A dataset of 6136 individuals over age 50 whocompleted a computerized cognitive testing battery for mild impairment was used. 5108 individuals had a primary diagnosis of one of 81 neurological or psychiatric conditions ("neurological condition" group), and 1028 were cognitively healthy (control group). Four ML algorithms (Naive Bayes, multi-layer perceptron [MLP] neural network, support vector machine [SVM], decision tree) were compared in separating the neurological condition and control group using 65 raw cognitive scores, age and education. For the best performing algorithm, confidence in individual classifications and separation among diagnostic groups were evaluated. Results: TheMLP neural network approach best separated neurological condition and control group (area under the curve [AUC]=0.89; compare with AUC=0.70 for a single global cognitive score). In a test sample, maximum confidence (from MLP output node activity) was assigned in 76% of individuals classified as "neurological condition", with 85% of these correctly classified. Maximum confidence was assigned in 71% of individuals classified as cognitively healthy, with 87% of these correctly classified. MLP performed well in separating the control group from mild cognitive impairment (MCI) (AUC=0.82), dementia (AUC=0.89), Parkinson's disease (AUC=0.91), multiple sclerosis (AUC=0.94), stroke (AUC=0.89), and traumatic brain injury (AUC=0.93), and showed good separation between amnestic MCI and Alzheimer's disease (AUC=0.75). Conclusions: MLalgorithms may greatly improve the accuracy and efficiency of the neuropsychologist or cognitive expert in integrating thewealth of computerized cognitive scores with other clinical information in assessing older adults. Future work should evaluate ML algorithms that incorporate other clinical data as well as their potential utility in predicting conversion to MCI or dementia.