The Cognitive, Age, Functioning, and Apolipoprotein E4 (CAFE) Scorecard to Predict the Development of Alzheimer′s Disease: A White-Box Approach

Abstract

Objective: This study aimed to bridge the gap between the costliness and complexity of diagnosing Alzheimer′s disease by developing a scoring system with interpretable machine learning to predict the risk of Alzheimer′s using obtainable variables to promote accessibility and early detection. Participants and Methods: We analyzed 713 participants with normal cognition or mild cognitive impairment from the Alzheimer′s Disease Neuroimaging Initiative. We integrated cognitive test scores from various domains, informant-reported daily functioning, APOE genotype, and demographics to generate the scorecards using the FasterRisk algorithm. Results: Various combinations of 5 features were selected to generate ten scorecards with a test area under the curve ranging from 0.867 to 0.893. The best performance scorecard generated the following point assignments: age < 76 (-2 points); no APOE ϵ4 alleles (-3 points); Rey Auditory Verbal Learning Test <= 36 items (4 points); Logical Memory delayed recall <= 3 items (5 points); and Functional Assessment Questionnaire <= 2 (-5 points). The probable Alzheimer′s development risk was 4.3% for a score of -10, 31.5% for a score of -3, 50% for a score of -1, 76.3% for a score of 1, and greater than 95% for a score of > 6. Conclusions: Our findings highlight the potential of these interpretable scorecards to predict the likelihood of developing Alzheimer′s disease using obtainable information, allowing for applicability across diverse healthcare environments. While our initial scope centers on Alzheimer′s disease, the foundation we have established paves the way for similar methodologies to be applied to other types of dementia. Keywords: Alzheimer′s disease; Machine learning; Cognition; Apolipoprotein ϵ4