Derivation and validation of a prediction model for primary and recurrent

Objective: Methods: We conducted a retrospective cohort study of HCT recipients (2010-2023) at a single institution. The cohort was randomly split into derivation (70%) and internal validation (30%) sets. We compared logistic regression with backward elimination by Akaike information criterion (AIC), random forest, and LASSO regularization approaches. Candidate predictors included demographics, clinical variables, laboratory values, and medication exposures. Model discrimination was assessed using the C-statistic, and calibration by observed vs. predicted proportions across risk deciles. Results: Among 2,725 HCT recipients, 252 (9.3%) developed CDI within one year, and 22 (8.7% of primary CDI cases) developed recurrence. For primary CDI, the backward elimination model performed best, including five predictors: receipt of cephalosporins (OR 1.46; 95% CI, 1.02-2.11), sulfonamides (OR 1.78; 95% CI, 1.04-3.05), penicillins (OR 1.4; 95% CI, 0.96-2.02), autologous transplant (OR 0.39; 95% CI, 0.22-0.66), and insurance type (Medicare: OR 30.2; 95% CI, 16.9-53.6; Medicaid: OR 15.2; 95% CI 6.5-35.4). This model showed good discrimination (C-statistic: 0.81 in both derivation and validation cohorts) with adequate overall calibration. For recurrent CDI, elevated white blood cell count at primary diagnosis was the only independent predictor (OR 1.16; 95% CI, 1.06-1.27), with modest discrimination (C-statistic: 0.73 derivation, 0.70 validation). Conclusion: We derived and internally validated prediction models for CDI in HCT recipients which could facilitate targeted preventive interventions in this high-risk population.