A data-driven dialogue system to enhance medical training with focus on comorbidity constructs

Medical education recently started to utilize large clinical datasets to provide, to medical students, data-driven tools to practice with clinical scenarios. In this work we explain a dialogue system that was designed to enhance medical education training. The objective of the dialogue system is to assist medical students and residents practice recognizing patient comorbidities, in an interactive way, and learn about the risk of each comorbidity construct for hospital death, inpatient septicemia, as well as the length of stay. The dialogue system provides to the learner functionality to initiate learning sessions to develop comorbidity profiles. Every time the learner adds a new comorbidity, the system recalculates conditional probabilities and provides new recommendations to the learner. A system backend can be used to parameterize the comorbidities that will qualify to the recommender. These parameters control the number of recommended comorbidities to the learner, and therefore the difficulty of the system. The study was completed with a large dataset with more than 11 million hospital admissions of Medicare patients, in the United States. The algorithm contains a customized Bayesian approach with feedback loops: The trainee initiates a scenario by selecting demographics and a principal diagnosis, then adds chronic and hospital-acquired conditions to see a summary of the attributable risk in each phase. Conditional probabilities are calculated and updated for the three hospital outcomes every time the learner adds new comorbidities per recommender suggestion. To examine the perceived usefulness of the dialogue system, a scenario-based evaluation was completed by 18 medical residents, who evaluated the system navigation, usefulness, validity of information and system features. The ĝ€?FOR VERIFICATION>system features' dimension was found to be the one ranked with the highest score (4.15/5), while the "information validity"received the lowest score (3.77/5). The composite variable about the participants' opinion on data-driven education was found to be high (4.27/5), demonstrating a strong demand for next generation medical training applications.