Performance of four computer-coded verbal autopsy methods for cause of death assignment compared with physician coding on 24,000 deaths in low- and middle-income countries

Table 2 Description of testing on multiple computer-coded verbal autopsy methods and datasets

Dataset	Training/testing cases	Number of diagnostic indicators
		King-Lu	Open-source random forest	Open-source tariff method s	InterVA-4
China	1100 / 400	48	48	48	N/A
Institute for Health Metrics and Evaluation	1100 / 400	96	96	96	N/A
Million Death Study	1100 / 400	89	89	89	N/A
	1100 / 1100	89	89	89	N/A
	6100 / 6100^a	89	89	89	245
Agincourt	1100 / 400	104	104	104	245^b
	1100 / 1100	104	104	104	245
	2900 / 2900	104	104	104	245
Matlab	1100 / 400	224	224	224	245
	1100 / 1100	224	224	224	245
	1600 / 1600	224	224	224	245

Only the numbers of test cases are applicable for the InterVA-4 analyses, as this method does not require any training cases. Additionally, InterVA-4 requires the input of 245 diagnostic indicators, however as many of these were not available in the given datasets, the number of useable variables was lower than 245. ^aThe MDS dataset used for InterVA-4 contained 552 cases, in which we extracted additional InterVA-4 indicators from the narratives. ^bEach CCVA method ran 30 resamples for each training/testing split within each dataset, except InterVA-4, which used the following number of re-samples: 1 for MDS data; 8, 7, 6 for Agincourt data splits of 400, 1100, and 2900 test cases; and 10, 10, 10 for Matlab data splits of 400, 1100, and 1600 test cases, respectively.

ISSN: 1741-7015