Table 5.

AD/NC classification comparison of our proposed method (iMSF) and missing value estimation methods (Zero, KNN, SVD and EM) in terms of accuracy, sensitivity and specificity when the training percentage varies from 1/2 to 3/4 as well as leave-one-out (LOO). In this experiment, we used the full multi-source data including MRI, PET, proteomics and CSF with 383 subjects in total.

Accuracy	Training Size	iMSF	Zero	EM	KNN	SVD
	50.0%	0.8658	0.8571	0.8737	0.8615	0.8404
	66.7%	0.889	0.8667	0.8814	0.8733	0.8494
	75.0%	0.8848	0.8662	0.8887	0.8701	0.8554
	LOO	0.9082	0.8671	0.8987	0.8797	0.8481
Sensitivity	Training Size	iMSF	Zero	EM	KNN	SVD
	50.0%	0.8552	0.8808	0.8879	0.8842	0.8539
	66.7%	0.8706	0.8838	0.8785	0.8818	0.8468
	75.0%	0.8667	0.8793	0.885	0.8849	0.8523
	LOO	0.8951	0.8889	0.8951	0.8827	0.8519
Specificity	Training Size	iMSF	Zero	EM	KNN	SVD
	50.0%	0.879	0.8354	0.861	0.8407	0.8279
	66.7%	0.9142	0.8481	0.8844	0.8636	0.8523
	75.0%	0.9102	0.8551	0.8942	0.8569	0.8593
	LOO	0.9351	0.8442	0.9026	0.8766	0.8442