Data Source and Study Cohorts
This study was conducted using the French National Health Insurance database (Système National des Données de Santé [SNDS])[17,18] and followed the REporting of studies Conducted using Observational Routinely-collected health Data guidelines. The database covers 98.8% of the population living in France (approximately 66 million inhabitants) and contains exhaustive data on all reimbursements for health-related expenditures, including dispensed drugs with date of dispensation. Information about all hospitalizations in a public or private hospital is also provided, including diagnoses (using International Classification of Diseases 10th revision [ICD]-10 codes) and expensive drugs prescribed during hospital stays. Long-term diseases (including cancers) are recorded, with diagnoses encoded according to ICD-10, because they give entitlement to 100% health insurance coverage.
We previously constructed a national cohort including every new patient receiving a systemic treatment for metastatic melanoma in France between June 2015 and December 2017 (n = 4725). The population selection process has been detailed previously. We selected 2 cohorts of patients with metastatic melanoma, differentiated by the first-line treatment received: an anti-PD-1 antibody cohort including 2605 patients receiving pembrolizumab or nivolumab and a targeted therapy cohort including 1527 patients receiving BRAF inhibitors (vemurafenib, dabrafenib) alone or combined with MEK inhibitors (cobimetinib, trametinib).
The study was approved by the French Data Protection Agency (CNIL, DR-2016-384).
A time frame for antibiotic courses received in the 3 months prior to the initiation of the first-line antimelanoma treatment was chosen for the main analysis, so as to correspond to the previous literature regarding antibiotic exposure and ICI efficacy among patients with melanoma. Antibiotic exposure was ascertained from dispensations occurring outside the hospital and was assumed to last 7 days, except for certain packages of tetracyclines used for acne and rosacea (60 days), single-dose fosfomycin (1 day), and benzathine penicillin G (14 days).
In a first sensitivity analysis, we narrowed the time frame for antibiotic exposure to 1 month prior to the first-line antimelanoma treatment, as this timing appeared as more detrimental for ICI efficacy.[9,11] In a second sensitivity analysis, antibiotic exposure was restricted to antibiotics having a higher impact on the gut microbiota, namely fluoroquinolones and broad-spectrum β-lactams (including second- and third-generation cephalosporins and penicillin associated with β-lactamase inhibitors).[21–24] These can be considered as the backbone for testing the hypothesis of impaired efficacy of ICI as a result of antibiotic-induced gut dysbiosis. Patients receiving high-impact antibiotics were compared with patients who did not receive antibiotics. A third sensitivity analysis was performed because administration of antibiotics is not recorded during hospital stays in the SNDS database. Patients with an infection identified in hospital discharge coding, who could have been misclassified in the "no antibiotics" group if they had received antibiotics only during their hospital stay, were excluded.
Two outcomes were used. Overall survival (OS) was estimated from the initiation of the first treatment line to the date of death or until censoring on December 31, 2017. The duration of the first treatment line, estimated from time-to-treatment discontinuation (TTD), was defined as the date of the start of the first treatment line to the date of treatment discontinuation. The date of the start of the first treatment line was the date of the first dispensation of targeted therapy or infusion of anti-PD-1 antibody. The treatment line was considered as discontinued when another treatment line was initiated or 3 months after the last recorded dispensation if no subsequent treatment was initiated. TTD has been proposed as an efficacy endpoint for real-world evidence trials.[25,26]
Multiple covariates were used: age, sex, number and location of metastatic sites, previous surgery, and stereotactic or external beam radiotherapy. It seems as if comorbidities were: - cardiovascular or cerebrovascular disorders - diabetes - history of another cancer - chronic respiratory, renal, liver, or pancreatic disease etc. whereas cardiovascular disorders, cerebrovascular disorders, chronic respiratory disease, chronic renal disease, chronic liver disease, chronic pancreatic disease, etc. were all considered separately (see Supplementary Table 1, available online). The algorithm used to identify comorbidities in the SNDS database was based on the SNDS mapping tool. This consists in reference-coding algorithms developed to standardize the analysis of the morbidity burden from health-care utilization.[27,28] Long-term disease diagnoses, hospitalization discharge diagnoses, and medical procedures were sought in the 4 years prior to the antimelanoma treatment initiation. In addition, medications were screened for in the 12 months before the antimelanoma treatment initiation (4 years for certain immunosuppressive or immunomodulating agents).
Propensity Score and Pseudo-Populations
Overlap weighting based on the propensity score was used to create pseudo-populations, in which characteristics measured were balanced between patients exposed and nonexposed to antibiotics.[29,30] A multivariate logistic regression was performed to calculate the propensity score for antibiotic exposure using all the covariates with non-zero values collected before the antibiotic time frame. Treated patients were weighted by the probability of not receiving antibiotics (1−propensity score), and untreated patients were weighted by the probability of receiving antibiotics (propensity score).[31,32] Covariate balance was checked using standardized differences between the exposed and non-exposed groups, in the original populations and in pseudo-populations.
Kaplan-Meier curves and weighted log-rank tests were performed within pseudo-populations. To evaluate the impact of antibiotic exposure on OS and TTD, a propensity score–weighted Cox proportional hazards regression model was used to estimate weighted hazard ratios (wHRs) and 95% confidence intervals (CIs). Proportional hazards were assessed by plotting log(-log(survival)) vs the log of survival time for categorical covariates and scaled Schönfeld residuals vs survival time for continuous covariates. To study the prescription of antibiotics according to the progression of melanoma disease, the proportion of patients initiating an antibiotic course was plotted against time for patients receiving anti-PD-1 antibody or targeted therapy as a first treatment line. Because every individual living in France is included in the SNDS database until death, there was no loss to follow-up. Statistical significance was defined at an a priori value of .05. Statistical analyses were performed using R v3.6.0 software. All statistical tests were 2-sided.
J Natl Cancer Inst. 2022;114(5):686-694. © 2022 Oxford University Press