Discussion

In this large retrospective cohort study, we derived and internally validated prediction models for primary and recurrent CDI among the HCT population. Our primary CDI model identified key predictors, including receipt of cephalosporins and sulfonamides, insurance type, and transplant type, achieving good discrimination with a C-statistic of 0.81. We also developed a clinical prediction tool for recurrent CDI based on WBC count at primary CDI diagnosis, which demonstrated moderate discrimination with a C-statistic of 0.70–0.73.

We evaluated multiple modeling approaches to balance predictive performance with clinical interpretability. Traditional logistic regression with backward elimination offers transparency, which is important for the practical application of infection prevention and antimicrobial stewardship programs. LASSO regression provides regularization to reduce overfitting in high-dimensional clinical data, while random forest methods allow for modeling of potential nonlinear relationships. In our cohort, the backward-elimination model showed the most stable performance suggesting that parsimonious models may be better suited for clinical settings characterized by relatively low event rates.

We identified cephalosporin use as a significant predictor for primary CDI among HCT recipients which is consistent with prior research. ^{Reference Aldrete S del, Kraft, Magee, Chan, Hutcherson and Langston28,Reference Tariq, Furqan, Jamshed and Khanna29} However, our finding that sulfonamide use was associated with increased CDI risk contrasts with a previous study by Lavallée et al., which reported sulfonamides as potentially protective in allogenic HCT patients. ^{Reference Lavallée, Labbé, Talbot, Alonso, Marr and Cohen21} Notably, numerous studies in the general population have identified sulfonamides as a risk factor for both hospital-acquired and community-acquired CDI. ^{Reference Teng, Reveles, Obodozie-Ofoegbu and Frei30–Reference Slimings and Riley35} The protective effect of autologous transplant (OR 0.39) is consistent with previous studies showing lower CDI incidence in autologous compared to allogenic HCT recipients. ^{Reference Alonso, Treadway, Hanna, Huff, Neofytos and Carroll6,Reference Shah, McClellan, Flowers, Lonial, Khoury and Waller7,Reference Lavallée, Labbé, Talbot, Alonso, Marr and Cohen21,Reference Guddati, Kumar, Ahmed, Ali, Kumar and Hari36}

We observed a strong association between public insurance (Medicare/Medicaid) and risk of CDI. While one previous study found public insurance predictive of primary CDI among general hospitalized patients, ^{Reference Rafalko, Webster, Jacob, Kutzler and Goldstein17} the magnitude of association in our study was substantial. In post hoc analyses, replacing insurance status with age ≥65 years, chronic kidney disease, and area deprivation index reduced model performance. This suggests that insurance status captures additional risk factors beyond these measured variables. Medicare beneficiaries include a heterogeneous population, including younger patients with end-stage renal disease or disabilities, who may possess distinct comorbidity profiles. Therefore, insurance status likely functions as a proxy for unmeasured social determinants of health, geographic proximity to the transplant center, or disparities in healthcare access, rather than a direct causal factor. Insurance type may also reflect variation in pretransplant care, antibiotic exposure, and healthcare-seeking behaviors, all of which may influence CDI risk. These findings should therefore be interpreted cautiously, and identification of more granular social and healthcare-related contributors to CDI risk in transplant recipients remains an important area for future investigation.

For recurrent CDI, we identified elevated WBC count at primary CDI diagnosis as the only significant predictor, consistent with studies in the general hospitalized population. ^{Reference Larrainzar-Coghen, Rodriguez-Pardo, Puig-Asensio, Rodríguez, Ferrer and Bartolomé37,Reference Reveles, Mortensen, Koeller, Lawson, Pugh and Rumbellow38} To our knowledge, this is the first study identifying severity of leukocytosis as a predictor for recurrent CDI among HCT recipients. However, given the low number of recurrent CDI cases, use of this model should be interpreted with caution.

Our primary CDI prediction model showed discrimination comparable to or better than models developed for the general hospitalized population. ^{Reference Panchavati, Zelin, Garikipati, Pellegrini, Iqbal and Barnes14,Reference Oh, Makar, Fusco, McCaffrey, Rao and Ryan15,Reference Rafalko, Webster, Jacob, Kutzler and Goldstein17,Reference Aukes, Fireman, Lewis, Timbol, Hansen and Yu18} While overall calibration was adequate, the model tended to underestimate risk in lower-risk groups and overestimate risk in higher-risk groups. The observed sensitivity of 69% indicates that this tool represents an initial proof-of-concept and that additional refinement, potential incorporation of novel predictors, and external validation will be required to further improve performance characteristics. Ultimately, such models may support IDSA/SHEA-recommended preventive strategies by identifying high-risk patients for targeted antimicrobial stewardship and infection prevention measures. The model may also help identify patient subsets for enhanced surveillance or early intervention strategies but should not be used to guide treatment decisions without external validation.

Our recurrent CDI model showed modest discrimination, similar to existing models for the general population. ^{Reference Madden, Boone, Lee, Sifri and Petri16,Reference Ruzicka, Kondo, Fujimoto, Craig, Kim and Mikamo19,Reference D’Agostino, Collins, Pencina, Kean and Gorbach20,Reference Escobar, Turk, Ragins, Ha, Hoberman and LeVine39} The lower discrimination compared to our primary CDI model suggests that important predictors of recurrence such as subsequent antibiotic exposures and C. difficile ribotype may not be captured in current clinical data. Despite its limitations, the model could help identify patients who might benefit from extended CDI treatment courses or novel microbiome-based therapies or fecal microbiota transplantation to prevent recurrence.

Our clinical prediction tools have the potential to stratify patients into distinct CDI risk categories. This risk stratification can be used to guide targeted preventive interventions. For high-risk patients, a bundled approach could include enhanced infection control measures, antibiotic stewardship focused on high-risk antibiotics such as cephalosporins and sulfonamides, and prophylactic strategies such as oral vancomycin prophylaxis. Importantly, this prediction model should be viewed as a tool in development, rather than a platform ready for immediate clinical implementation. While the primary CDI model demonstrated good discrimination, further refinement and external validation are necessary before such models can reliably be incorporated into routine infection prevention practice or used to guide antimicrobial stewardship interventions.

Our study has several limitations. First, it was conducted at a single institution, with most patients identifying as White and having private insurance, which may limit the generalizability of the findings. However, this remains one of the largest, well-characterized cohorts of HCT recipients, spanning 13 years. Second, the strong association observed with insurance status requires further investigation to identify underlying causal factors. Third, changes in institutional CDI diagnostic testing criteria over time may have influenced our findings. Furthermore, we were unable to account for potential differences in CDI testing frequency across patient subgroups (eg, those with inflammatory bowel disease), which may have introduced detection bias, particularly during the PCR-only testing era when colonization could have been misclassified as infection. Although we excluded patients with CDI within 8 weeks prior to HCT, earlier CDI history may not have been captured and could confer residual risk, potentially leading to misclassification. Variation in CDI treatment strategies, including antimicrobial selection, duration of therapy, and use of adjunctive treatments, may influence recurrence risk and were not fully accounted for in this analysis. This heterogeneity may partially explain the modest discrimination of the recurrent CDI model and suggests that future studies should incorporate granular treatment data. Finally, our recurrent CDI model was limited by the small number of recurrent events (n = 22), which restricted our ability to identify multiple predictors and perform detailed calibration analyses.