Bayesian Double Feature Allocation for Phenotyping With Electronic Health Records

Abstract

We propose a categorical matrix factorization method to infer latent diseases from electronic health records (EHR) data in an unsupervised manner. A latent disease is defined as an unknown biological aberration that causes a set of common symptoms for a group of patients. The proposed approach is based on a novel double feature allocation model which simultaneously allocates features to the rows and the columns of a categorical matrix. Using a Bayesian approach, available prior information on known diseases greatly improves identifiability and interpretability of latent diseases. This includes known diagnoses for patients and known association of diseases with symptoms. We validate the proposed approach by simulation studies including misspecified models and comparison with sparse latent factor models. In the application to Chinese EHR data, we find interesting results, some of which agree with related clinical and medical knowledge.

Consensus Monte Carlo for Random Subsets Using Shared Anchors

“…Latent diseases X2 and X3 are associated with the same set of symptoms but with opposite signs. This interesting result is also found in Ni et al (2019) where X2 and X3 were identified as polycythemia and anemia, respectively. Each of X5, X6 and X8 also finds good correspondence in Ni et al (2019) as bacterial infection, viral infection and thrombocytopenia.…”

“…Unlike MNIST or the application to tumor heterogeneity, there is no ground truth or alternative implementation for posterior inference for the full data. Instead we compare the results with previous results by Ni et al (2019) who used a full MCMC implementation for a subset of 1000 patients from the same dataset. Some of our findings are consistent with the earlier results, which suggests a good approximation of the proposed CMC to full MCMC.…”

“…This is known as bi-clustering (Hartigan, 1972). Ni et al (2019) introduces a similar prior model for random row and column subsets, but now without the restrictions of a partition, i.e., random feature allocation on rows and columns simultaneously. Figure 1 illustrates…”

“…That is, A describes the patient-disease relationships and C describes symptom-disease relationships. We follow the same procedure in Ni et al (2019) to preprocess the data. Using the reference range for each test item, we discretize the data and define a symptom if the value of an item falls beyond the reference range.…”

“…One prevelant latent disease reported in Ni et al (2019) does not have a clear interpre- Table 2: Blood test items. Acronyms are given within parentheses.…”