JMIR Publications

ABSTRACT

Background:

Cohort studies contain rich clinical data across large and diverse patient populations that are a common source of observational data for clinical research. Because large scale cohort studies are both time and resource intensive, one alternative is to harmonize data from existing cohorts through multi-cohort studies. Given differences in variable encoding, however, accurate variable harmonization is difficult.

Objective:

We propose SONAR, a method for harmonizing variables across cohort studies, in order to facilitate multi-cohort studies.

Methods:

SONAR ensembles semantic learning from variable descriptions and distribution learning from study participant data. Our method learns an embedding vector for each variable and uses pairwise cosine similarity to score the similarity between variables. This approach was built off three NIH cohorts, including the Cardiovascular Health Study, the Multi-Ethnic Study of Atherosclerosis, and the Women’s Health Initiative. We also use gold-standard labels to further refine the embeddings in a supervised manner.

Results:

The method was evaluated using manually curated gold-standard labels from the three NIH cohorts. We evaluated both the intra-cohort and inter-cohort variable harmonization performance. The supervised SONAR method outperforms existing benchmark methods for almost all intra-cohort and inter-cohort comparisons using AUC and top k accuracy metrics. Notably, SONAR is able to significantly improve harmonization of concepts that are difficult for existing semantic methods to harmonize.

Conclusions:

SONAR achieves accurate variable harmonization within and between cohort studies by harnessing the complementary strengths of semantic learning and variable distribution learning.