RIDDLE: Race and ethnicity Imputation from Disease history with Deep LEarning
Table 4
DeepLIFT contribution score orderings for 10 most predictive ICD9 codes.
DeepLIFT scores were computed using separate test samples and models from ten k-fold cross validation experiments; scores were summed across experiments. DeepLIFT scores were produced for each pair of feature, and output (race and ethnicity) class; we list ten ICD9 codes with the highest ranges of scores—which correspond to discriminative ability. The feature-to-class contribution scores were used to construct orderings of race and ethnicity classes, for each feature. Scores were summed across all samples. Positive scores indicate favorable contribution to a class, zero scores indicate no contribution, and negative scores indicate discrimination against a class.