Dementia screening with machine learning methods

Machine learning algorithms were applied to an electronic patient database generated by the UC Irvine Alzheimer's clinic to learn the simplest and most accurate patient parameters that would discriminate 244 very mildly demented from 198 normally aging subjects. Attributes included age, sex and education plus responses to the Functional Activities Questionnaire, the Mini-Mental Status and Blessed Orientation, Memory and Concentration tests. The machine learning algorithms included decision tree learners (C4.5, CART), rule inducers (C4.5 Rules, FOCL) and naive Bayes. Stepwise logistic regression was used to compare results. The sample was randomly split into training and testing sets, and the results were validated over 30 runs. Although the Functional Activities Questionnaire has been used since 1980, the machine learning algorithms were the first to identify that a single attribute, measuring forgetfulness, equaled or exceeded the accuracy of any other scoring method of this test. Post hoc inspection of the odds ratios obtained by stepwise logistic regression confirmed this finding. The application of machine learning has identified an extremely simple, yet accurate screen for very mild dementia.