Distributed Multi-site Latent Class Analysis
- Description
- dMLCA illustration
The latent class analysis (LCA) model is a widely recognized and effective method for identifying disease subphenotypes. The emergence of clinical research networks enables multi-site data integration, providing more generalizable and reliable analysis. However, implementing LCA in a multi-site setting is challenging due to the prohibition of patient-level data sharing and the heterogeneity among different sites. The traditional divide-and-conquer strategy (e.g., meta-analysis) cannot be applied because the number and characteristics of latent classes learned on each site may be different and thus hard to combine. We develop a distributed multi-site latent class analysis (dMLCA) approach. dMLCA allows multiple sites to learn a common set of latent classes collaboratively without sharing patient-level data and can effectively account for the heterogeneity among different sites. In addition, dMLCA can incorporate patient-level covariates in class membership modeling. We showed the validity and strength of the dMLCA through simulation. We applied it to identify subphenotypes of multisystem inflammatory syndrome in children (MIS-C) using data from 864 patients from nine children's hospitals.
