It is my distinct honor and privilege to announce that we have another doctor in the community! Last week, Rupa Makadia successfully defended her dissertation, as part of her PhD from Rutgers University in the Department of Health Informatics, under the advising of Dr. Shankar Srinivasan. I was honored to participate on Rupa’s committee, alongside Dr. Frederick Coffman and @Christian_Reich . Rupa’s dissertation, “Development and evaluation of a machine learning algorithm to map medical conditions and procedures from real-world data”, was a successful demonstration of the power of integrating biomedical ontologies and data science to produce data-driven insights that can improve our ability to generate reliable real-world evidence, and a great example of research enabled by the OMOP common data model and ability to run analyses across a network of disparate databases. @rmakadia specifically tackled the problem of creating a structured ontology that relates diagnoses to the medical procedures used to 1) diagnosis and 2) treat those conditions. While colloquially we can all understand that ‘appendectomy’ is a surgical intervention used to treat ‘appendicitis’ and ‘mammography’ is used to diagnosis ‘breast cancer’, the reality has been that these relationships have never existed in any structured dataset but rather were buried in the unstructured freetext of medical textbooks or manually curated in disease-specific repositories…until now! Rupa used data from MarketScan, Optum, and Premier to infer this medical knowledge by exploring temporal co-occurrence in current clinical practice to derive a predictive model that could discriminate between known condition-procedure relationships and a large panel of >32k negative control pairs. The performance Rupa achieved, with AUC>0.95, was truly impressive, and performing external validation across 5 databases is something that few published predictive models (let alone work from a graduate student) ever achieve. In total, Rupa was been able to map corresponding procedures for >90% of conditions, and find conditions for >60% of the procedure codes we use in our routine work. This condition-procedure ontology should be a tremendous asset to OHDSI and the broader research community in allowing us new ways to explore patient-level data, design phenotypes, and construct features for large-scale modeling.
In her PhD defense, Rupa did a wonderful job of clearly and succinctly explaining the relevance of her work to the informatics field as an enabling technology to improve observational research, highlighting the results of model training and external validation across 5 databases, and illustrating the impact of the model in automatically learning a corpus of condition-procedure relationships. The committee was very impressed by the polish in her presentation, and main consistent feedback was strong encouragement to publish two peer-review publications about her new condition-procedure ontology and the characterization of the results, so that others in the research community can learn from and use her important contribution.
For those who are interested in hearing more about Rupa’s work, she will be presenting on today’s OHDSI community call (Tuesday, March 5, 12pmET-1pmET, details at: Weekly OHDSI Digest - 4Mar2019). I look forward to seeing an encore performance from @rmakadia and also am eager to hear feedback from the rest of the community of future collaborative research we can pursue given this exciting progress.