Abstract
INTRODUCTION: Receiving timely Alzheimer's disease (AD) diagnosis is often delayed due to long waitlists for specialists. Our study aimed to bridge the gap between the timeliness and complexity of diagnosing AD by developing a scoring system with interpretable machine learning using variables that are obtainable at integrated primary care settings. METHODS: We trained the model using 666 participants with normal cognition or mild cognitive impairment at baseline visit from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and externally validated the scorecard using 4,876 participants from the National Alzheimer's Coordinating Center (NACC). We integrated cognitive measures, daily functioning measured with Functional Assessment Questionnaire (FAQ), and demographics into FasterRisk algorithm. RESULTS: Combinations of 4 separate measures were selected to generate 10 scorecards, showing strong performance (area under the curve [AUC] = 0.868-0.892) in ADNI and remaining robust when externally validated in NACC (AUC = 0.795). The features were Category Animal ≤ 20 (2 points), Trail Making Test B ≤ 143 (-3 points), Logical Memory Delayed ≤ 3 (4 points), Logical Memory Delayed ≤ 8 (3 points), and FAQ ≤ 2 (-5 points). The probable AD risk increased correspondingly with higher total points: 7.4% (-8), 25.3% (-4), 50% (-1), 74.7% (2), and >90% (>6). We refer to this model as the (F)unctioning, (LA)nguage, (M)emory, and (E)xecutive functioning or FLAME scorecard. INTERPRETATION: Our findings highlight the potential to predict AD development using obtainable information, allowing for implementation into integrated primary care workflows to initiate early intervention. While our scope centers on AD, this established foundation paves the way for other types of dementia.