Table 2.

Performance of the deep learning models

	Accuracy	Sensitivity	Specificity	F1-score	MCC
MRI
ADNI test	0.834 ± 0.020	0.767 ± 0.036	0.889 ± 0.030	0.806 ± 0.024	0.666 ± 0.042
AIBL	0.870 ± 0.022	0.594 ± 0.119	0.924 ± 0.025	0.593 ± 0.088	0.520 ± 0.095
FHS	0.766 ± 0.064	0.901 ± 0.096	0.712 ± 0.123	0.692 ± 0.044	0.571 ± 0.056
NACC	0.818 ± 0.033	0.764 ± 0.031	0.849 ± 0.052	0.757 ± 0.033	0.613 ± 0.059
Non-imaging
ADNI test	0.957 ± 0.010	0.924 ± 0.019	0.983 ± 0.032	0.951 ± 0.010	0.915 ± 0.020
AIBL	0.915 ± 0.022	0.872 ± 0.037	0.923 ± 0.034	0.772 ± 0.035	0.731 ± 0.035
FHS	0.760 ± 0.042	0.517 ± 0.043	0.842 ± 0.068	0.512 ± 0.026	0.367 ± 0.053
NACC	0.854 ± 0.021	0.881 ± 0.013	0.838 ± 0.041	0.817 ± 0.019	0.703 ± 0.033
Fusion
ADNI test	0.968 ± 0.014	0.957 ± 0.014	0.977 ± 0.031	0.965 ± 0.014	0.937 ± 0.026
AIBL	0.932 ± 0.031	0.877 ± 0.032	0.943 ± 0.042	0.814 ± 0.054	0.780 ± 0.059
FHS	0.792 ± 0.039	0.742 ± 0.185	0.808 ± 0.082	0.633 ± 0.076	0.517 ± 0.098
NACC	0.852 ± 0.037	0.924 ± 0.025	0.810 ± 0.068	0.824 ± 0.032	0.714 ± 0.053

Three models were constructed for explicit performance comparison. The MRI model predicted Alzheimer’s disease status based upon imaging features derived from the patch-wise trained FCN. The non-imaging model consisted of an MLP that processed non-imaging clinical variables (age, gender, MMSE). The fusion model appended the clinical variables used by the MRI model to the MLP portion of the non-imaging model in order to form a multimodal imaging/non-imaging input. Accuracy, sensitivity, specificity, F1-score, and Matthew’s correlation coefficient (MCC) are demonstrated for each. The fusion model was found to outperform the other models in nearly all metrics in each of the four datasets. Of interest, however, we noted that the performance of the MRI model and the non-imaging model still displayed higher specificity and sensitivity than many of the human neurologists, all of whom used the full suite of available data sources to arrive at an impression.