It is truly my honor and privilege to get to announce that we have a new doctor within our community. Yesterday, Dr. Joel Swerdel (@jswerdel) successfully defended his dissertation to his committee, and will soon be formally awarded his PhD in Epidemiology from Rutgers University.
Dr. Swerdel’s dissertation is titled: “Predictive modeling of incident heart failure in subjects with newly diagnosed atrial fibrillation”, and was under the direction of Dr. George Rhoads. In his dissertation presentation, Joel did a masterful job of presenting the clinical rationale of why comorbid atrial fibrillation and heart failure presents such a public health concern, and how developing a patient-level prediction model could improve clinical practice. He then showcased important aspects of the impressive volume of independent research that he designed and executed: he trained LASSO regression models using >10,000 covariates to predict ‘among patients who are newly diagnosed with atrial fibrillation, which patients will go on to develop heart failure in the future?’, and he systematically varied the target population (exploring patients <65yo and >65yo), the outcome (looking at all heart failure, and also partitioning out heart failure with preserved ejection fraction and heart failure with reduced ejection fraction), and the time-at-risk (looking at 1yr and 1-3yr time-horizon). For each model, he performed internal validation with a hold-out sample, and external validation by applying the model to a different database (e.g. training a model on Optum and applying to the IBM MDCR). The results were remarkably encouraging; Joel’s models achieved AUC>0.70 in most contexts and had strong face validity. Also of particular novelty, there have been no predictive models published in this particular topic, but quite a wealth of literature on the associated ‘risk factors’: Joel trained a model using only these published risk factors, and showed that limiting to this small set of ‘known’ baseline covariates achieves much lower discrimination (further demonstrating that the large-scale approach to model fitting can yield substantial gains even if we don’t have a priori hypotheses for all predictors in advance).
Joel’s work also represents an important milestone for OHDSI. To my knowledge, Joel’s dissertation is the first PhD fully supported by the OHDSI ecosystem. Joel used three different observational databases, IBM CCAE, IBM MDCR, and Optum Clinformatics, which have all been converted to the OMOP Common Data Model v5. He trained and externally validated his models using the OHDSI PatientLevelPrediction package. As Joel commented in his defense, the OHDSI tools made it possible to complete a large number of modeling design combinations efficiently, which helped support him completing such a large quantity of robust scientific inquiry in a relatively short amount of time.
Dr. Rhoads and his other committee members, including Dr. Elizabeth Marshall and Dr. Will Kostis, were keen to see Joel move beyond the dissertation and encouraged submission of manuscripts to high-impact cardiology journals, recognizing the novelty and importance of the research that Joel has completed. Given the strong performance of the model, Joel was also encouraged to consider ways to get this model into real clinical practice, either as a payer tool or embedded into EHR clinical decision support or directly translated into a patient-facing risk calculator. When Joel’s work get out from the Rutgers offices and into the public domain, I think there’s a real chance that Joel’s work could positively impact the lives of many patients around the world.