Validation of Clinical Risk Models for Clostridioides difficile-Attributable Outcomes

• Published in: Antimicrobial agents and chemotherapy Vol. 66; no. 7; p. e0067622
• Main Authors: Madden, Gregory R., Petri, William A., Costa, Deiziane V. S., Warren, Cirle A., Ma, Jennie Z., Sifri, Costi D.
• Format: Journal Article
• Language: English
• Published: 1752 N St., N.W., Washington, DC American Society for Microbiology 19-07-2022
• Subjects: Epidemiology; Epidemiology and Surveillance; Clostridioides difficile; prediction model; disease severity; outcome; risk model
• ISSN: 0066-4804; 1098-6596
• DOI: 10.1128/aac.00676-22

Clostridioides difficile is the leading health care-associated pathogen, leading to substantial morbidity and mortality; however, there is no widely accepted model to predict C. difficile infection severity. Most currently available models perform poorly or were calibrated to predict outcomes that are not clinically relevant. We sought to validate six of the leading risk models (Age Treatment Leukocyte Albumin Serum Creatinine (ATLAS), C. difficile Disease (CDD), Zar, Hensgens, Shivashankar, and C. difficile Severity Score (CDSS)), guideline severity criteria, and PCR cycle threshold for predicting C. difficile-attributable severe outcomes (inpatient mortality, colectomy/ileostomy, or intensive care due to sepsis). Models were calculated using electronic data available within ±48 h of diagnosis (unavailable laboratory measurements assigned zero points), calibrated using a large retrospective cohort of 3,327 inpatient infections spanning 10 years, and compared using receiver operating characteristic (ROC) and precision-recall curves. ATLAS achieved the highest area under the ROC curve (AuROC) of 0.781, significantly better than the next best performing model (Zar 0.745; 95% confidence interval of AuROC difference 0.0094–0.6222; P = 0.008), and highest area under the precision-recall curve of 0.232. Current IDSA/SHEA severity criteria demonstrated moderate performance (AuROC 0.738) and PCR cycle threshold performed the worst (0.531). The overall predictive value for all models was low, with a maximum positive predictive value of 37.9% (ATLAS cutoff ≥9). No clinical model performed well on external validation, but ATLAS did outperform other models for predicting clinically relevant C. difficile-attributable outcomes at diagnosis. Novel markers should be pursued to augment or replace underperforming clinical-only models.