Derivation and Validation of a Clinical Prediction Rule for Complications of Clostridium difficile Infection Using a Multicenter Prospective Cohort

BACKGROUND: Clostridium difficileinfection (CDI) outbreaks were associated with increase in unfavorable outcomes. Identifying and predicting risk of developing complications (cCDI) early in the course of illness could improve clinical decision-making. We developed and validated a prediction rule for cCDI. METHODS: Adult inpatients with confirmed CDI in 10 Canadian hospitals were enrolled and followed for 90 days. Data within 48h of CDI diagnosis were collected: demographics, underlying illnesses, past medical and drug history, clinical signs, blood tests, and strain ribotype. cCDI was defined as one or more of: colonic perforation, toxic megacolon, colectomy, need of vasopressors, ICU admission due to CDI, or if CDI contributed to 30-day death. Predictors’ selection was supported by experts’ opinion suggesting 17 clinical criteria. Cross-validation technique was used (2:1 ratio) and multivariable logistic regression for predictive modeling in the derivation subset. The optimal model was assessed by area under ROC curve (AUC) and prediction error (PE). A predictive score was built by assigning points proportional to adjusted risk estimates. RESULTS: Among 1380 patients enrolled, 1050 were used for predictive modeling (median age 70 years and one-third infected by ribotype 027 strains). Cases were split into training (n = 700) and validation sets (n = 350). A cCDI occured in 8% and 6.6% respectively. The optimal model with a PE of 5% and an AUC of 0.84 in the validation set included WCC (< 4, 12–19.9, or ≥20 × 10(9)/L), BUN≥11 mmol/L, serum albumin <25 g/L, heart rate > 90/minute, and respiratory rate >20/minute. A predictive score of min 0 and max 13 points was derived. A score ≥7 points was associated with 70% cases of cCDI, showed 68% sensitivity (95% CI, 55–80) in the derivation set and 70% (51–88) in the validation set, a specificity of 73% (69–76) and 76% (72–81) respectively, 17% PPV (9–25), and 97% NPV (95–99) in both sets. CONCLUSION: Using a large multicenter prospective cohort and robust modeling approach, we derived a predictive score that included easily available measures at the bedside. The score showed acceptable performance. Further validation is needed on cohorts with different characterstics (non-outbreak setting, higher rate of cCDI). Other approaches such as combination of biomarkers could be more predictive of cCDI. DISCLOSURES: J. Powis, Merck: Grant Investigator, Research grant; GSK: Grant Investigator, Research grant; Roche: Grant Investigator, Research grant; Synthetic Biologicals: Investigator, Research grant