P0193 - Predictive Machine Learning Models for Recurrent C. difficile Infection in Inflammatory Bowel Disease Patients Based on First-Episode CDI Treatment

Raseen Tariq, MBBS, Ankita Sethi, BA, Shivaram Poigai Arunachalam, PhD, DBA, Darrell S.. Pardi, MD, MS, William A. Faubion, MD, Sahil Khanna, MBBS, MS
Mayo Clinic, Rochester, MN

Introduction: Recurrent C difficile infection (rCDI) is common in patients with Inflammatory bowel disease (IBD). Data regarding the impact of primary CDI treatment on rCDI is lacking. We aimed to study the differential effect of antibiotics to treat CDI on risk of rCDI using Machine Learning (ML).

Methods: Data were divided into three cohorts based on the initial CDI treatment. Models for each cohort were trained using structured data to predict CDI recurrence based on the treatment of choice. CDI recurrence was defined as a positive stool test and antibiotics prescribed within 8 weeks after the initial episode, with relevant data from the preceding three months being considered. The final models were developed utilizing the XGBoost classifier, which was trained on the top 25 features identified through Recursive Feature Elimination. Performance of the models was evaluated using accuracy, the Area under the Receiver Operating Characteristic curve (AuROC), precision, sensitivity (recall) and specificity.

Results: A total of 2573 patients with IBD and their first episode of CDI were divided into four treatment cohorts: vancomycin (1377 patients), metronidazole (322 patients), fidaxomicin (67 patients), and vancomycin with metronidazole combination (581 patients). Among those, 297, 85, 17 and 210 had rCDI respectively. Using the top 25 features we trained three XGBoost model. The overall model performances were comparable across cohorts, with AuROC of 0.83 (vancomycin), 0.80 (fidaxomicin), and 0.81 (metronidazole) and lower for cohort receiving both vancomycin and metronidazole (0.66) (Table).

Further analysis of the top ranked features revealed distinct patterns for each treatment cohort. In the vancomycin cohort, the top three features for recurrence were male gender, ulcerative colitis, and azathioprine use within three months prior to the first CDI. For metronidazole, hospitalization at the time of CDI, female gender, and azathioprine use within three months were the top features. In the fidaxomicin cohort, mesalamine and cephalosporin use within three months prior to CDI, along with the presence of Crohn's disease, were most significant features of recurrence.

Discussion: Our ML models have identified distinct features within each treatment cohort that can serve as valuable clinical decision tools, enabling physicians to assess differential recurrence risks and personalize CDI treatments based on individual patient profiles.



Disclosures:


Raseen Tariq, MBBS, Ankita Sethi, BA, Shivaram Poigai Arunachalam, PhD, DBA, Darrell S.. Pardi, MD, MS, William A. Faubion, MD, Sahil Khanna, MBBS, MS. P0193 - Predictive Machine Learning Models for Recurrent C. difficile Infection in Inflammatory Bowel Disease Patients Based on First-Episode CDI Treatment, ACG 2023 Annual Scientific Meeting Abstracts. Vancouver, BC, Canada: American College of Gastroenterology.