Fig. 4: MCI classification using machine learning underscoring the role of neural features in performance.

a Model performance when using only the data from ADCS-ADL-MCI survey score. Note that the model, while good at detecting healthy controls, does not have good sensitivity to MCI. b Combining behavioral metrics from tasks with survey responses did not improve the model performance. c When using survey data as well as both neural and behavioral features from tasks, the model performance was starkly improved. In (a-c) Left) Model prediction raw scores for each group (x-axis). Shaded gray areas demonstrate the utility of adapting thresholds to capture an “Inconclusive” population. Middle) ROC curve with AUC shown. The diagonal indicates chance level. Right) The confusion matrix when model outputs are binarized.