A 5-min Cognitive Task With Deep Learning Accurately Detects Early Alzheimer's Disease

Abstract

Introduction: The goal of this study was to investigate and compare the classification performance of machine learning with behavioral data from standard neuropsychological tests, a cognitive task, or both. Methods: A neuropsychological battery and a simple 5-min cognitive task were administered to eight individuals with mild cognitive impairment (MCI), eight individuals with mild Alzheimer's disease (AD), and 41 demographically match controls (CN). A fully connected multilayer perceptron (MLP) network and four supervised traditional machine learning algorithms were used. Results: Traditional machine learning algorithms achieved similar classification performances with neuropsychological or cognitive data. MLP outperformed traditional algorithms with the cognitive data (either alone or together with neuropsychological data), but not neuropsychological data. In particularly, MLP with a combination of summarized scores from neuropsychological tests and the cognitive task achieved ~90% sensitivity and ~90% specificity. Applying the models to an independent dataset, in which the participants were demographically different from the ones in the main dataset, a high specificity was maintained (100%), but the sensitivity was dropped to 66.67%. Discussion: Deep learning with data from specific cognitive task(s) holds promise for assisting in the early diagnosis of Alzheimer's disease, but future work with a large and diverse sample is necessary to validate and to improve this approach.

Keywords: Alzheimer's disease; artificial neural networks; inhibition of return; machine learning; neuropsychological test.

Figure 1

The cognitive task [spatial inhibition of return (IOR)] experiment paradigm. Within each trial, there were three sequentially presented visual stimuli—two cues (solid red square) and one target (solid green square)—with a blank screen in between. The three stimuli were presented serially. The two cue stimuli could appear in any of the three locations (left, middle, right), whereas the target stimuli could only appear in one of the two locations (left or right, but not the middle). Subjects were instructed to respond to the target (solid green square) by pressing one of two buttons in the right hand to indicate whether the target was presented at the left or right location (with the index finger or the middle finger). The two cues were presented 200 ms each, with a 250 ms break in between. The second cue was followed by another 250 ms break before the onset of the target, which was presented for 850 ms. The next trial started 750 ms after the offset of the target stimulus. Subjects had to respond within the 1.6 s time-window (before the onset of next trial). There were five conditions based on the relationship of the locations in which the three stimuli were presented: aaa, in which the two cues and the target were presented at the same location; abb, in which the second cue and the target were presented at the same location, and the first cue was presented at a different location; aba, in which the first cue and the target were presented at the same location, and the second cue was presented at a different location; aab, in which the two cues were presented at the same location, and the target was presented at a different location; abc, in which the two cues and the target were presented at three different locations. The behavioral data from the study team can be found elsewhere (Jiang et al.), which includes detailed data from each individual subject that can be downloaded by other teams to test with their approaches. Note: ms, millisecond.

Figure 2

The ROC curves for the best classifiers selected by the highest sensitivity for each dataset with traditional machine learning algorithms and with MLP. See Table 2, Supplementary Tables 1-3 for the specific algorithms and parameters used for these “best” classifiers (shown in bold and italics font). (A) Traditional machine learning algorithms with all features and PCA (without and with SMOTE over-sampling). (B) Traditional machine learning algorithms with features selection (without and with SMOTE over-sampling). (C) The ROC curves for each dataset with MLP using the demographically comparable dataset. The AUC score is shown in the legend box.