Naive Bayes classifiers for verbal autopsies: comparison to physician-based classification for 21,000 child and adult deaths

Table 6 Mean CSMF accuracy (and 95 % uncertainty intervals) on three datasets for 35 train/test iterations

Study (training/testing sample size)^a	NBC	OTM	InterVA-4^b	Median all three classifiers
MDS	0.88	0.57	0.71	0.71
(11,000/555)	(0.87,0.88)	(0.56,0.57)	(0.69,0.73)	0.71
Agincourt	0.87	0.42	0.66	0.66
(2,300/2,300)	(0.87,0.88)	(0.42,0.43)	(0.63,0.68)	0.66
Matlab	0.92	0.57	0.65	0.65
(1,000/1,000)	(0.92,0.93)	(0.56,0.58)	(0.62,0.67)	0.65
Median all three datasets	0.88	0.57	0.66	0.66

CSMF cause-specific mortality fraction, NBC naïve Bayes classifier, OTM open-source Tariff Method, VA verbal autopsy, MDS Million Death Study
^aTraining/testing sample size, with no training required for InterVA-4
^bInterVA-4 was evaluated on a testing data set of 50 randomly selected records out of 555 records, in each of the 35 iterations

ISSN: 1741-7015