We constructed a longitudinal cohort of adults (≥18 years) undergoing C. difficile testing within the Duke University Health System, including 3 hospitals and their affiliated outpatient clinics, between 15 March 2020 and 31 December 2023. The study period corresponded with the onset of 2-step testing within our health system. Analysis was performed from January 2024 to May 2024. Duke University Health System’s institutional review board reviewed the study design and determined it to be exempt with a waiver of consent for data collection as the study was entirely retrospective.

Antibiotic receipt was electronically queried across our entire health system, capturing inpatient administrations and outpatient prescriptions during the study period. For censoring purposes, the date of the last encounter was defined as the last in-person clinic or hospital encounter occurring within 90 days of the index C. difficile test. Clostridioides difficile status was defined by 2-step testing using NAAT (Cepheid Xpert) with reflex to toxin enzyme-linked immunosorbent assay (TechLab Quik Chek Complete) when positive. Per clinical microbiology laboratory protocols, samples were only tested if an unformed stool sample was received. NAAT⁺/toxin⁻ results are accompanied by the following text: “Likely represents colonization with alternative cause of symptoms. Treatment generally not indicated unless clinical signs of severe or fulminant C. difficile disease. Consider Infectious Disease consult in questionable cases.” Fulminant disease was defined as cases requiring rectally administered vancomycin or colectomy.

Three separate time-to-event analyses were conducted: (1) assessing risk of progression to toxin positivity, (2) describing treatment rates among NAAT⁺/toxin⁻ individuals, and (3) exploring the association between presumptive treatment and subsequent toxin positivity and fulminant disease outcomes. In each model, follow-up time began with C. difficile testing and continued up to 90 days from the index test. Subjects were considered censored at either the date of their last in-person clinical encounter or the date of death if either fell within 90 days of the index test. More than 1 test-and-follow-up cycle could be included for each subject; thus, robust confidence intervals (CIs) and models appropriate for recurrent events (Prentice-Williams-Peterson gap-time and frailty models) were employed. Models were constructed using the survival and cmprsk packages in R (version 4.1.2, https://cran.r.project.org) [6, 7]. The proportionality of hazards was assessed for each model using a combination of graphical (log-log plots) and test-based (ZPH test with Schoenfeld residuals) methods. Variables violating the proportionality assumption were treated as stratification variables. Key potential confounders were treated as time-varying covariates, including age, chronic kidney disease, cancer, hematopoietic stem cell transplant history, 2 or more prior C. difficile episodes, and high-risk antibiotic receipt in the 30 days before the episode [8]. High-risk antibiotics were defined as fluoroquinolones, third- and fourth-generation cephalosporins, or clindamycin [9]. Proton pump inhibitor use was not included, partly because prior local cohorts did not find a significant association and partly because over-the-counter availability limits the utility of electronic queries for prescription data [10]. Competing risks were accounted for using the cmprsk package in analysis 2. A significance level of .05 was used for analysis 1. Analyses 2 and 3 were descriptive or exploratory, and thus P values were not reported.

Data from the full cohort of tested subjects compared the risk of subsequent toxin⁺ C. difficile episodes among NAAT⁺/toxin⁻ subjects relative to NAAT⁻ and relative to recurrent disease among NAAT⁺/toxin⁺ subjects. Since the assumptions inherent to various recurrent event models are not easily testable, we conducted analysis 1 in triplicate—using the Prentice-Williams-Peterson gap-time method, the frailty model method, and a traditional Cox proportional hazard model (limited to the first event only) to ensure stability. To adjust for immortality time bias (retesting is generally not permitted within 14 days of a prior NAAT⁺/toxin⁺ test within our healthcare system), the follow-up period for NAAT⁺/toxin⁺ subjects was adjusted to begin 14 days after the last positive test. To investigate the potential influence of informative censoring, we conducted sensitivity analyses in which censored cases were first treated as worst-case scenario (eg, progression to toxin positivity at the time of censoring) and best-case scenario (eg, remained disease free at the time of censoring). Potential interaction terms between antibiotic receipt and C. difficile status were explored and included when appropriate based on likelihood ratio testing.

To assess rates of treatment among NAAT⁺/toxin⁻ subjects—including either presumptive treatment or treatment for subsequent toxin⁺ episode—we constructed cumulative incidence curves for the subcohort of subjects whose index C. difficile test was NAAT⁺/toxin⁻. Receipt of long-term C. difficile suppressive therapy and all-cause mortality were included as competing risks.

To explore the association between presumptive treatment of NAAT⁺/toxin⁻ subjects and later toxin positivity, we again constructed time-to-event models within the cohort of NAAT⁺/toxin⁻ subjects while excluding those placed on long-term suppression (defined as continuation of a C. difficile–active therapy for a duration longer than the guideline-recommended maximum of 14 days). To adjust for immortality time bias introduced by presumptive treatment, time at risk for presumptively treated subjects was set to begin at completion of the presumptive treatment course. Treatment with rectally instilled vancomycin (given in cases of severe C. difficile with ileus) or colectomy were selected as electronically queryable indicators of fulminant disease.

Unadjusted Kaplan-Meier curves plotting C. difficile toxin⁺ episode-free survival are shown in Figure 2. As there was evidence of interaction between C. difficile status and antibiotic receipt (P < .001 by likelihood ratio test), we created a new variable representing each combination of C. difficile status and antibiotic receipt to allow for easier effect interpretation with proportional hazard modeling. Hazard ratios adjusted for age, select comorbidities, and prior C. difficile events are presented in Table 2. Relative to NAAT⁻ subjects who did not receive antibiotics, NAAT⁺/toxin⁻ status was associated with a higher risk of progression to toxin positivity (adjusted hazard ratio [aHR], 5.06 [95% CI, 3.61–7.10]). The hazard associated with NAAT⁺/toxin⁻ status was numerically lower than that associated with initial NAAT⁺/toxin⁺ status (aHR, 23.47 [95% CI, 17.61–31.28]). Across all statuses, high-risk antibiotic receipt increased the risk of progression to C. difficile toxin positivity, including a particularly high risk among NAAT⁺/toxin⁻ subjects receiving high-risk antibiotics (aHR, 15.71 [95% CI, 9.85–25.06]).

The general effects of C. difficile NAAT⁺/toxin⁻ status and antibiotic receipt remained stable across all 3 time-to-event models (Prentice-Williams-Peterson gap-time, frailty model, and a traditional Cox proportional hazard model limited to first events only; Supplementary Table 1). Though not the specific focus of this study, the hazard associated with initial NAAT⁺/toxin⁺ status remained significant but varied in magnitude across models, likely because of smaller numbers within this subgroup. Similarly, C. difficile NAAT⁺/toxin⁻ status and antibiotic receipt remained significantly associated with later progression to toxin positivity across best- and worst-case censoring assumptions, though the magnitude of risk was sensitive to the assumption that all censored subjects developed disease (Supplementary Table 2).

Figure 3 presents a cumulative incidence plot for clinical outcomes among NAAT⁺/toxin⁻ subjects. Presumptive treatment was prescribed for 28% of episodes (95% CI, 27%–30%) within 10 days of testing, increasing slightly to 33% (95% CI, 31%–35%) by day 90 (Supplementary Table 3). A small proportion of NAAT⁺/toxin⁻ subjects received long-term suppression (n = 24 [1.0%]). Progression to toxin positivity occurred in 5% (95% CI, 4%–6%) of NAAT⁺/toxin⁻ episodes within 90 days of their index test.

A total of 5 toxin⁺ episodes occurred following 690 NAAT⁺/toxin⁻ events receiving presumptive treatment (90-day Kaplan-Meier event rate: 2.8% [95% CI, 0–5.8%]; Figure 4A); 83 toxin⁺ episodes occurred following 1570 NAAT⁺/toxin⁻ events not receiving presumptive treatment (90-day Kaplan-Meier event rate: 7.6% [95% CI, 5.9%–8.2%]). Using rectal vancomycin receipt or colectomy as indicators of severe C. difficile, just 2 colectomies occurred after presumptive treatment (90-day Kaplan-Meier event rate: 0.7% [95% CI, 0–1.7%]); 5 colectomies plus 1 instance of rectal vancomycin receipt occurred among untreated subjects (90-day Kaplan-Meier event rate: 0.6% [95% CI, 0–1.1%]; Figure 4B). One colectomy from the untreated group was for colonic necrosis occurring 6 weeks after index test and presumed to be C. difficile related though stool remained NAAT⁺/toxin⁻ at the time of colectomy. The other 6 colectomies were all for unrelated indications (4 for diverticular disease, 1 for colon cancer, 1 for ulcerative colitis). While smaller event rates limited the ability to include all potential confounders, presumptive treatment appeared to be associated with a lower hazard of progression to toxin positivity (aHR, 0.13 [95% CI, .05–.31]) but did not appear to be associated with any reduction in risk of treatment for severe C. difficile disease or colectomy (aHR, 1.93 [95% CI, .50–7.45]) (Table 3). Event rates in analysis 3 were too small to adequately assess additional factors for those most likely to benefit from treatment.