Teaching data science fundamentals through realistic synthetic clinical cardiovascular data

Abstract

ObjectiveOur goal was to create a synthetic dataset and curricular materials to assist in teaching fundamentals of translational data science.\n\nMaterials and MethodsA literature review was conducted to extract current cardiovascular risk score logic, data elements, and population characteristics. Then, clinical data elements in the models were pulled from clinical data and transformed to the Observational Medical Outcomes Partnership (OMOP) common data model; genetic data elements were added based on population rates. A hybrid Bayesian network was used to create synthetic data from the logical elements of the risk scores and the underlying population frequencies of the clinical data.\n\nResultsA synthetic dataset of 446,000 patients was created. A two-day curriculum was created based on this synthetic data with exploratory data analysis and machine learning components. The curriculum was offered on two separate occasions; the two groups of learners were given the curriculum and data, and results were tallied, summarized, and compared. Students ability to complete the challenge was mixed; more experienced students achieved a range of 70%-85% in balanced accuracy, but many others did not perform better than the baseline model.\n\nDiscussionOverall, students enjoyed the course and dataset, but some struggled to consistently apply machine learning techniques. The curriculum, data set, techniques for generation, and results are available for others to use for their own training.\n\nConclusionA realistic synthetic data with clinical and genetic components helps students learn issues in cardiovascular risk scoring, practice data science skills, and compete in a challenge to improve identification of risk.