Fig. 6: Diagnostic code prediction using the MIMIC-III and Quebec CHD datasets.

a Proposed MixEHR+RNN framework for predicting medical code from the MIMIC-III data. b Medical code-prediction accuracy. We evaluated the prediction accuracy of the 42 common Clinical Classification Software (CCS) medical codes comparing MixEHR with Doctor AI and GRAM. For each medical code, the accuracy was calculated by the true positives among the top 20 predicted codes divided by the total number of positives. The dots are individual accuracies for each code predicted by each method. c Proposed MixEHR+RNN framework for predicting the ICD code on Quebec CHD database. d Prediction accuracies in terms of area under the precision–recall curve (AUPRC) and area under the ROC curve (AUROC) over all of the 1098 3-digit ICD codes. The center line, bounds, and whiskers of the boxplots are median, first and third quartiles, and outlier, respectively. Wilcoxon signed-rank one-sided tests were performed to calculate the p-values with the alternative hypothesis set to that AUCs from MixEHR+RNN are greater than the AUCs from the baseline RNN.