(SLA-065) Applications of Machine Learning in Epidemiologic Research Utilizing Administrative Claims and/or Structured Electronic Health Record Data

Background: Machine Learning (ML) is increasingly applied in epidemiologic research that utilize real world data (RWD). Systematic understanding of types of ML algorithms employed, how they are trained and evaluated, and their comparative performance could inform future applications improvement.

Objectives: To characterize the application of ML in epidemiologic research that utilized administrative claims and structured electronic health record (EHR) data.

Methods: A TLR was conducted using structured searches of Embase and MEDLINE databases on March 19, 2025. The search strategy incorporated Emtree terms, MeSH headings, and keywords related to claims and structured EHR data, ML and algorithms, predictive modeling, clustering, classification, and disease identification. Two reviewers first independently screened the titles and abstracts, then screened the full text using predefined eligibility criteria and alignment or resolution by a third reviewer. Eligible articles were extracted using a standardized form that captured data sources, patient characteristics, outcomes and disease measures, ML and traditional statistical approaches used, including elements of the ML algorithms including feature, training, validation and evaluation, tools and libraries used, findings, and limitations.

Results: Of the 120 publications identified, 42 unique studies were included in this TLR. Twenty-five used predictive modeling, 13 used classification modeling, and 4 studies included both predictive and classification modeling. A total of 23 studies used administrative claims data while 9 studies used structured EHR and EMR data; 9 studies used both claims and EHR/EMR data and 1 study used both claims and survey data. Many ML approaches were used in the studies reviewed and most (n=29) used multiple ML approaches. Random forest and random forest variations were the most common ML approaches used to predict outcomes and identify diseases (n=32). The second most common ML method used was boosting, with 16 studies employing gradient boosting methodologies. Among studies that utilized multiple ML approaches, the best performing models for prediction were boosting models, and the top performing classification model was logistic regression.

Conclusions: We found that while RF and gradient boosting performed best for prediction tasks using RWD, logistic regression performed better for classification. It is not clear to what extent researchers’ expertise and familiarity with different approaches drove performance, but newer techniques did not always outperform traditional ones.