AI Tool Shows Promise for Antimicrobial Stewardship, C Diff Prevention in Hospitals

Use of an AI-guided infection prevention bundle may support Clostridioides difficile infection (CDI) prevention and antimicrobial stewardship in hospital settings, according to findings from a recent study.1

The prospective, single-center quality improvement study was conducted at Michigan Medicine, the academic medical center affiliated with the University of Michigan, and compared patient outcomes pre- and post-implementation of a previously validated institution-specific AI model for CDI risk prediction. Results showed that although use of the AI bundle was not associated with a significant reduction in CDI incidence, it was linked to substantial reductions in CDI-associated antimicrobial use.1

“Numerous artificial intelligence models have been developed to predict CDI risk in hospitalized patients with the goal of identifying high-risk patients who may potentially benefit from targeted infection prevention resources,” Jenna Wiens, PhD, an associate professor of computer science and engineering at the University of Michigan, Ann Arbor, and colleagues wrote.1 “Although these models have shown promise in prospective validation and in identifying patients prior to colonization, their association with patient outcomes in clinical care remains unknown.”

According to the US Centers for Disease Control and Prevention, Clostridioides difficile, a bacterium that causes diarrhea and colitis, is estimated to cause almost half a million infections in the United States each year.2 Because Clostridioides difficile can live on people's skin and is resistant to alcohol-based hand sanitizers, effective CDI prevention often includes soap-and-water handwashing. Receipt of certain antimicrobials is also associated with increased susceptibility to CDI.1

Applying these practices at scale can be challenging due to resource constraints at most hospitals, necessitating targeted prevention efforts for patients at the greatest risk of acquiring CDI.1

To evaluate the association between an AI-guided infection prevention bundle and CDI incidence in a hospital setting, investigators evaluated adult inpatient hospitalizations before and after the integration of a previously validated institution-specific AI model into clinical workflows at Michigan Medicine. The CDI Risk Score is calculated by an L2-regularized logistic regression model with time-varying parameters and leverages routinely collected EHR data to estimate the daily risk of a patient acquiring CDI in the remainder of their hospitalization.1

The pre-AI period was from September 2021 to August 2022, and the post-AI period was from January 2023, to December 2023. Of note, investigators excluded the 4-month implementation rollout period from September 2022 to December 2022 from the analysis.1

The primary outcome was CDI incidence rate. Secondary outcomes included antimicrobial use and qualitative assessments of bundle implementation.1

Pre- and post-AI samples included 39,046 and 40,515 hospitalizations, respectively. In the pre-AI group, the majority of patients were female (55.4%) and White (78.8%) with a median (IQR) age of 58 (IQR, 36 to 70) years. In the post-AI group, the majority of patients were female (55.7%) and White (79.6%) with a median age of 58 (IQR, 37 to 70) years. Investigators called attention to significant differences in age, COVID-19– related hospitalizations, and recent hospitalization history between the groups.1

There were 127 CDI cases for 255 017 patient-days in the pre-AI period and 148 CDI cases for 261 892 patient-days in the post-AI period. After adjusting for differences in clinical characteristics, the incidence CDI rate per 10 000 patient-days was 5.76 (95% CI, 4.87 to 6.69) in the pre-AI period and 5.65 (95% CI, 4.78 to 6.56) in the post-AI period (absolute difference, −0.11; 95% CI, −1.43 to 1.18; P = .85).1

Investigators called attention to significant reductions in the use of ampicillin sulbactam (−2.82; 95% CI, −4.59 to −1.03; P = .03), piperacillin-tazobactam (−9.64; 95% CI, −12.93 to −6.28; P <.001), and clindamycin (−1.04; 95% CI, −1.60 to −0.47; P = .03), measured in antibiotic days per 1000 days present. Of note, reductions were more prominent in high-risk hospitalizations alerted by the AI tool.1

Using qualitative assessments via semistructured interviews and field observations, investigators found health care staff’s experiences with AI-guided workflows varied. Results showed an overall low adherence rate for the enhanced hand hygiene sign, while pharmacists consistently engaged with the alerts and contacted the patients’ care team based on the alerts.1

“Despite limitations of our single-center, quasi-experimental design, the insights gained from this study are broadly applicable to other institutions seeking to use AI models to improve clinical outcomes,” investigators concluded.1 “Future research should focus on developing novel strategies of ongoing education and stakeholder engagement to overcome implementation barriers, exploring the scalability of these efforts across diverse health care settings, identifying optimal alert thresholds that balance sensitivity and alert burden, and conducting randomized clinical trials to assess their effect on clinical outcomes.”

References

1. Tang S, Shepard S, Clark R, et al. Guiding Clostridioides difficile Infection Prevention Efforts in a Hospital Setting With AI. JAMA Netw Open. 2025;8(6):e2515213. doi:10.1001/jamanetworkopen.2025.15213

2. US Centers for Disease Control and Prevention. About C. diff. December 18, 2024. Accessed June 12, 2025. https://www.cdc.gov/c-diff/about/index.html