About 20-40% of patients with incident Clostridium difficileinfection (iCDI) will have a recurrence (rCDI). Identifying high risk patients to prevent rCDI is an important clinical and public health objective. Existing rCDI predictive models have emphasized use of administrative and manually abstracted data. The objective of our study was to develop methods to harness detailed data from a comprehensive, commercially available electronic medical record (EMR) to improve rCDI prediction.
A comprehensive EMR retrospective cohort study dataset using Kaiser Permanente Northern California data from 2007-2014. Data included C. difficiletest results, demographic and administrative data, inpatient bed histories, outpatient encounters, microbiologic results, medications (e.g., antibiotics, proton pump inhibitors), laboratory tests, and vital signs. Multiple algorithms that permitted rule-based approaches were used to define iCDI, refractory CDI, and rCDI and included antibiotic treatment windows for iCDI and rCDI to create an episode-based structure. The final episode-based dataset is suitable for rapid analysis using traditional, Bayesian, and machine learning analytic methods.
We identified 22,259 adults with iCDI, of which 3,178 met criteria for rCDI (defined as a positive test occurring 4 days after the end of a treatment window to 84 days). Of the iCDI cases, 63% have vital signs (temperature, heart rate, respiratory rate, oxygen saturation). A LAPS2 (a composite severity of illness composite score) could be generated for 72% of patients. For the entire cohort, data were available for previous antibiotics use, GI surgery, hospital encounters, comorbidity burden, and immunosuppression status (based on diagnosis and medication use).
Availability of detailed EMR data permits development of larger and comprehensive CDI datasets than previously available. These datasets permit a more flexible modeling structure that permit rapid variation of specific eligibility criteria (e.g., time frame for recurrence vs. refractory status) for modeling and sensitivity analyses. Leveraging large and generalizable EMR data combined with novel analytic techniques could improve upon existing prediction models for rCDI.
Merck & Co. Inc.:
J. Greene, Merck & Co. Inc.: Grant Investigator and Investigator , Research grant
J. C. Laguardia, Merck & Co. Inc.: Grant Investigator and Investigator , Research grant
N. Cossrow, Merck & Co. Inc.: Employee and Shareholder , Salary
E. R. Dubberke, rebiotix: Consultant and Investigator , Consulting fee and Research support
sanofi-pasteur: Grant Investigator , Research grant
pfizer: Consultant , Consulting fee
Merck: Consultant and Investigator , Consulting fee and Research support
See more of: Poster Abstract Session