Table 4.

MCI/NC classification comparison of the ensemble methods (Voting, Uniform and Learned) and the other methods (Best and Average) in terms of accuracy, sensitivity and specificity when the training percentage varies from 1/2 to 3/4. In this experiment, we used the full multi-source dataset including MRI, PET, proteomics and CSF, from 608 subjects in total.

Accuracy	Training Size	Voting	Uniform	Learned	Best	Average
	50.0%	0.8832	0.8754	0.8845	0.8872	0.8504
	66.7%	0.9105	0.8865	0.8967	0.9033	0.8591
	75.0%	0.9026	0.8821	0.893	0.8927	0.8573
Sensitivity	Training Size	Voting	Uniform	Learned	Best	Average
	50.0%	0.7829	0.4446	0.5051	0.6228	0.3698
	66.7%	0.8393	0.5119	0.582	0.6922	0.414
	75.0%	0.846	0.5031	0.5639	0.7162	0.4139
Specificity	Training Size	Voting	Uniform	Learned	Best	Average
	50.0%	0.9121	0.995	0.9901	0.9955	0.9837
	66.7%	0.9321	0.9938	0.9883	0.9956	0.9855
	75.0%	0.9212	0.9925	0.9888	0.9982	0.985