Screening for Early-Stage Alzheimer's Disease Using Optimized Feature Sets and Machine Learning

Background: Detecting early-stage Alzheimer's disease in clinical practice is difficult due to a lack of efficient and easily administered cognitive assessments that are sensitive to very mild impairment, a likely contributor to the high rate of undetected dementia. Objective: We aim to identify groups of cognitive assessment features optimized for detecting mild impairment that may be used to improve routine screening. We also compare the efficacy of classifying impairment using either a two-class (impaired versus non-impaired) or three-class using the Clinical Dementia Rating (CDR 0 versus CDR 0.5 versus CDR 1) approach. Methods: Supervised feature selection methods generated groups of cognitive measurements targeting impairment defined at CDR 0.5 and above. Random forest classifiers then generated predictions of impairment for each group using highly stochastic cross-validation, with group outputs examined using general linear models. Results: The strategy of combining impairment levels for two-class classification resulted in significantly higher sensitivities and negative predictive values, two metrics useful in clinical screening, compared to the three-class approach. Four features (delayed WAIS Logical Memory, trail-making, patient and informant memory questions), totaling about 15 minutes of testing time (~30 minutes with delay), enabled classification sensitivity of 94.53% (88.43% positive predictive value, PPV). The addition of four more features significantly increased sensitivity to 95.18% (88.77% PPV) when added to the model as a second classifier. Conclusion: The high detection rate paired with the minimal assessment time of the four identified features may act as an effective starting point for developing screening protocols targeting cognitive impairment defined at CDR 0.5 and above.