Fig. 7: Lab results imputation using MIMIC-III dataset.

a Workflow to impute lab results. Step 1: We modeled lab tests, lab test results, and non-lab EHR data (i.e., ICD, notes, prescription, treatment) to infer the patient topic mixture. Step 2: For a test patient, we masked each of his observed lab test result t, and inferred his topic mixture. Step 3: We then found k = 25 patients who have the lab test results t observed and exhibit the most similar topic mixture to the test patient. We then took the average of lab result values over the k patients as the prediction of the lab result value for the test patient \(j^{\prime}\). Steps 1–3 were repeated to evaluate every observed lab test in every test patient. b We compared the imputation accuracy between MixEHR and CF-RBM. We generated the cumulative density function (CDF) of accuracy as well as the boxplot distributions (inset) for each method. The center line, bounds, and whiskers of the boxplots are median, first and third quartiles, and outlier, respectively. In both cases, MixEHR significantly outperformed CF-RBM based on KS test (p < 1.15e-5) and Wilcoxon signed-rank one-sided test (p < 0.00013). c CF-RBM versus MixEHR scatterplot in terms of imputation accuracy.