ICD-10 code embedding for patient matching: a first step toward applications in causal inference from hospital discharge data

Abstract We present WEMatch, a Word2vec Enhanced ICD-10 Matching algorithm, for patient matching based on ICD-10 codes and make a first step toward evaluating its potential use in causal inference contexts by comparing it to several propensity score–based patient matching approaches. We use hospital discharge data from the Insel Group in the canton of Bern, Switzerland, from 2014 to 2023, including ICD-10 codes, age, sex, and BMI for 386 842 hospitalizations. In the null-effect matched cohort design, patients are matched to a randomly selected “treatment group” (selected from 1 of 4 selected primary ICD-10 diagnoses) using 4 algorithms: WEMatch and 3 propensity score matching (PSM)–based models. Performance is assessed by comparing the balance of key variables, hazard ratios of mortality, and length of hospital stay between the treatment and matched control groups. WEMatch and PSM methods perform similarly for 3 out of 4 tested ICD-10 codes. For one ICD-10 code, WEMatch outperforms PSM methods, because the patient population is especially diverse. This study establishes a robust framework for systematically comparing patient matching algorithms. The groundwork is set for applying WEMatch to clinically relevant questions. By distributing this as an R package, we promote its integration into clinical research, potentially supporting future causal analyses using observational data.