Antimicrobial exposure is associated with decreased survival in triple-negative breast cancer

a, b Overall survival (OS) and breast cancer-specific survival (BCS) estimates for any antimicrobial exposure. c, d OS and BCS estimates for total antimicrobial exposure, visualized into quartiles. e, f OS and BCS estimates for unique antimicrobial exposure, visualized into quartiles. Inset numbers represent the number of patients still at risk of death at each time point. Patients could move to increasing exposure categories over time.

Associations of tumor and patient characteristics with survival

Using marginal structural models (MSMs) to estimate the associations of individual covariates with mortality, we found, for all three exposure definitions, that higher cancer stage and undergoing unilateral mastectomy were associated with decreased OS and BCS (for any antimicrobial exposure: OS for cancer stage III versus I, 4.02 (2.42–6.69); OS for unilateral mastectomy versus lumpectomy, 1.60 (1.10–2.34); BCS for cancer stage III versus I, 6.33 (3.28–12.21); BCS for unilateral mastectomy versus lumpectomy, 1.90 (1.22–2.96); results were similar for total and unique antimicrobial use.

Sensitivity analysis

To address whether patients who are sicker at baseline or become frailer during treatment may have inferior outcomes related to their clinical performance status and not due to greater antimicrobial exposure, we considered disease severity, defined by receipt of intravenous antimicrobials and ICD10 codes associated with acute illness (sepsis, severe sepsis, and systemic inflammatory response syndrome) 90 days prior to or following antimicrobial prescriptions (in n = 54 patients), as an additional covariate in the MSM model. We did not observe a substantial change in OS or BCS estimates for any exposure definition (for any antimicrobial exposure: OS 1.45 (0.93–2.28), BCS 1.40 (0.84–2.34); for total exposures: OS 1.05 (1.02–1.08), BCS 1.05 (1.02–1.09); for unique exposures: OS 1.16 (1.11–1.21), BCS 1.17 (1.12–1.23)) (Supplemental Table 6).

To evaluate whether the observed mortality impact is ALC-specific, we compared the standard adjusted unweighted Cox regression model to the MSM. When accounting for the effect of ALC in the MSM, we observed a decrease in the HRs for both OS and BCS (OS for unweighted Cox 1.54 (0.98–2.49) versus OS for MSM 1.46 (0.93–2.29); BCS for unweighted Cox 1.47 (0.89–2.45) versus BCS for MSM 1.39 (0.84–2.32)), suggesting that ALC is associated with both outcomes. When accounting similarly for the effect of ANC, we observed no change in the HRs for OS or BCS (OS for unweighted Cox 1.58 (1.01–2.49) versus OS for MSM 1.58 (1.01–2.48); BCS for unweighted Cox 1.53 (0.92–2.54) versus BCS for MSM 1.52 (0.92–2.53)), suggesting ANC is not along the causal pathway from antimicrobial exposure to increased mortality (Supplemental Table 7).

Associations of antimicrobial exposure with survival over time

Given that patients with TNBC have an elevated risk of recurrence in the first 2–5 years post-diagnosis, we used landmark analysis to determine whether antimicrobial use was significantly associated with mortality at yearly intervals following diagnosis. This analysis considers the impact of ongoing exposure to antimicrobials after completion of breast cancer treatment on the subset of women who remain event-free but at risk of recurrence at each time point. For those who met the cumulative exposure definitions associated with inferior survival, we found evidence of a strong and sustained by-year association through year 3 post-diagnosis that then decreased at years 4 and 5 (Fig. 3).

We studied long-term OS and BCS associated with antimicrobial exposure in 772 TNBC patients treated with curative intent from 2000–2014 at two institutions representing community and academic practice. Consistent with the study’s hypothesis, we found that cumulative antimicrobial exposure after diagnosis of early-stage TNBC was associated with inferior OS and BCS. This association was related to time-varying ALC levels, was sustained in strength through year 3 post-diagnosis, and was independent of severe illness. To our knowledge, this is the first study to report an association of antimicrobial exposure during treatment with breast cancer outcomes and to explore the interaction of antimicrobials with ALC, which is prognostic in TNBC19. If confirmed, these findings may inform antimicrobial prescribing practices during TNBC treatment and follow-up.

We previously reported that lower ALC after TNBC diagnosis was associated with inferior OS and BCS, that the strength of this association increased with time, and that higher baseline TIL density was associated with higher peripheral lymphocyte count during treatment19, suggesting peripheral lymphocyte count may be a biomarker of tumor-directed immunity. Higher TIL density in pretreatment breast cancer biopsies is also associated with a better response to neoadjuvant chemotherapy2 and improved DFS and OS in patients with node-positive TNBC38, which has the most aggressive biology among breast cancer subtypes in terms of early recurrence and worst prognosis4, and disproportionately affects women of African American and Hispanic race and ethnicity and of lower socioeconomic status5. A recent epidemiological study showed impaired survival in breast cancer patients receiving antibiotics prior to diagnosis and treatment without ICIs, and most significantly in the month preceding diagnosis29, which lends support to the hypothesis that antimicrobial exposure may impair baseline tumor-directed immunity and perhaps TIL levels. Such at-risk patients, however, can only be identified retrospectively. In an exploratory, limited subset analysis, we confirmed that TIL levels at diagnosis were associated with a pathologic complete response to chemotherapy, as others have shown2, but were not associated with subsequent receipt of antimicrobials and antimicrobial exposure was not associated with pathologic complete response, though the current study’s sample size was limited due to a small number of historical samples available for TIL analysis and not sufficiently powered to detect these associations. Taken together, these findings suggest that antimicrobial exposure both before and during treatment is an immune-modulating risk factor for TNBC mortality, but that the relationship between antimicrobials, circulating and tumor-directed lymphocytes, and response to treatment is complex and dynamic. The impact of ICIs on these relationships will be an intriguing area of future study.

While lymphopenia may initially occur during receipt of cytotoxic chemotherapy and predispose patients to opportunistic infections39 and inferior outcomes19, emerging data suggest antimicrobial exposure may sustain impaired peripheral immunity secondary to gut microbiota disruption. This altered immunity may contribute to both cancer development40 and progression41. An active area of focus is the augmentation of host antitumor immunity: immunotherapy increases the rate of pathologic complete response, which is correlated with improved survival, when added to neoadjuvant chemotherapy7 and is also effective in heavily pre-treated TNBC13. Breast cancer intratumoral vaccination strategies are also being studied (ClinicalTrials.gov42 identifiers: NCT02018458, NCT01703754, and NCT02423902). The hypothesis that gut microbiome dysbiosis impairs the efficacy of ICIs is increasingly relevant for early-stage TNBC patients, given the 2021 United States Food and Drug Administration approval of the ICI pembrolizumab with neoadjuvant chemotherapy based on the KEYNOTE-522 trial7. The present findings demonstrate an association of increasing antimicrobial exposure with inferior TNBC outcomes related to lymphopenia in a population treated before ICI approval in breast cancer and suggest that antimicrobial exposure also impairs the efficacy of traditional cytotoxic chemotherapies. They also offer evidence that lymphopenia may be secondary to antimicrobial-mediated gut microbiome dysbiosis.

This study has limitations that warrant consideration. Since this is a retrospective, observational study using CCR data linked to EMR data collected during the routine course of care, we cannot infer causality and could not collect corollary microbiome samples from the analytic sample of nearly 800 TNBC patients treated since 2000. TIL scores at diagnosis were available for a small subset of patients enrolled in a clinical trial37, but were not available after chemotherapy and antimicrobial exposure for most patients. Some data are missing, including prescriptions and laboratory results obtained outside of the studied healthcare systems. Because prophylactic antimicrobials are rarely given in general medical practice and are not given in breast cancer, treatment was assumed to be for clinical infection, and we assumed that patients took medications as prescribed. We could not evaluate non-adherence and did not evaluate ICD codes associated with infection proximal to antimicrobial prescriptions, as prior studies have shown limited accuracy in clinical recording, especially of secondary diagnoses (e.g., infection as a secondary condition diagnosed in a physician office visit focused on breast cancer treatment)43. Most infections in adults are treated empirically (i.e., without culture-based evidence of infection with a specific organism) and are cured in the outpatient setting. Thus, we did not evaluate for chart documentation of infection with specific organisms nor development of antimicrobial resistance, as <1% of our sample was treated in the inpatient setting with intravenous antimicrobials where organisms and their antibiotic sensitivities are typically identified. We have also made several assumptions in statistical modeling. To address the consistency assumption, we evaluated three well-defined antimicrobial exposure definitions and separately evaluated mortality associations with each. We assumed that exposed and unexposed individuals had equivalent distributions of mortality predictors but acknowledge that there remain unmeasured confounders that may impact exposure and mortality associations, for which we cannot formally test given the observational nature of the study.

Although it is possible that excluded patients without recorded blood counts were healthier, the finding of a survival association with ALC but not ANC suggests a lymphocyte-specific observation, even though patients with both low ALC and ANC values were more likely to be exposed to antimicrobials. Similarly, while sicker patients may receive more antimicrobials and have inferior survival unrelated to an antimicrobial effect, we found a comparable BCS association after controlling for severe illness that continued for three years post-diagnosis, suggesting a sustained, illness-independent effect. As all patients had non-metastatic disease at diagnosis and were treated with curative intent, the longitudinal nature of this survival association suggests ongoing antimicrobial exposure after completion of treatment impacts the risk of recurrence related to the host immune response to residual disease, which parallels the clinical timeframe of greatest risk in these women. Given that we could not profile gut taxonomy or T-cell receptor phenotypes, we could not evaluate the relationship between antimicrobial exposure and gut microbiome composition or TIL subset evolution during chemotherapy, nor the relationship between circulating and tumor-associated lymphocytes through treatment. Consistent with our previous report44, we observed again here that unilateral mastectomy was associated with inferior survival compared to bilateral mastectomy or lumpectomy with radiation; this likely reflects unmeasured confounders, given equivalent survival after mastectomy versus breast-conserving therapy in randomized trials45,46, and not an antimicrobial-mediated survival effect. A small portion of patients did not have a documented surgical procedure in our analyzed dataset; as surgery is the standard of care for all early-stage TNBCs, it is likely that these patients had surgeries performed outside of California, which would not be captured by the CCR, or that they had comorbidities or other factors precluding surgery. The BMI of this study sample is lower, and neighborhood socioeconomic status is higher than national averages, which reflects a local demographic and may limit the findings’ generalizability.

This study’s limitations are balanced by its notable strengths, including innovative study design with a large sample of TNBC patients treated in different healthcare settings; integration of EMR and CCR data which enables a rich set of covariates; near-complete mortality capture through CCR protocols; and comprehensive statistical modeling of the impact of both ALC and antimicrobial exposure on survival outcomes over time. Furthermore, these robust, clinically relevant findings support the design of future prospective studies that will collect corollary microbiome and TIL data.