P0469 - Informing fellow education through automated extraction of clinical experiences from the electronic health record

Background: Graduate medical education (GME) strives to produce physicians that provide safe and efficient care. Despite this mission, the relationship between clinical experiences during training and clinical competence is poorly understood. As competency based medical education becomes realized, resulting educational structures will necessitate alignment of patient care experiences with standardized goals. Improved patient attribution using clinical informatics tools can facilitate quantification of these experiences and inform GME outcomes of measuring clinical performance and quality measures. We previously developed and validated the TRAinee Interactive Learning System (TRAILS) which gathers the patients and diagnoses that trainees encounter using predictive algorithms and electronic health record (EHR) data. Our hypothesis is that this system, which was originally designed for pediatric residents, can be used to describe clinical experiences of pediatric critical care medicine (PCCM) fellows.

Methods: We completed an IRB-approved retrospective study of 19 PCCM fellows from July 2017 to July 2018 in a 60-bed PICU. EHR elements (e.g. time-in-chart, notes, and orders) were used to identify shifts and clinical encounters. Encounters were classified as fellow provider-patient interactions (PPI) using previously-validated models. Clinical shifts, encounters, and PPI characteristics, including patient diagnoses are reported by post-graduate year (PGY).

Results: There was no statistical difference in shift counts and durations across PGY (overall median [IQR] shift count: 137 [123, 145]; mean ± SD duration (hrs): 14.1±6.9). Shift duration was distributed bimodally with peaks ~ 12 and 26 hours. Patient interaction counts per fellow (median [IQR]: 1314 [1104, 1400]) were not different across PGY (ANOVA F=0.19, p=0.83). Encounters classified as PPI (median count: 505 [428, 583]) demonstrated significantly higher total time-in-chart (961±692 vs 180±159 seconds) and more orders placed (8.9±15.0 vs 0.1±1.2) than those not classified as a PPI. The most common billing diagnoses were “GERD”, “Acute Respiratory Failure”, “Acidosis”, “Acute Kidney Injury”, and “Tachycardia”. Among the most common diagnoses, exposure varied significantly by PGY year.

Conclusion: Clinical experiences among PCCM fellows can be extracted using EHR data and predictive models, minimizing recall bias associated with manual entry. This system can be used to prospectively identify relative differences in experience to track and assure sufficient exposure in PCCM trainees. Future studies will examine exposure variation among fellowship programs and practice groups, potential associations between exposure and competence, as well as target GME outcomes and quality measures.