Association between Broad-Spectrum Antibiotic Use and Risk of Early-Onset Colorectal Cancer.

Introduction
Colorectal cancer (CRC) is the fourth most common cancer and the second
leading cause of cancer-related deaths in the US. 1 The incidence of CRC has been declining for
adults aged 50 and older since mid-1980’s2 in part due to screening.3 In contrast, the incidence of early-onset CRC (eoCRC)
diagnosed under age 50 has been increasing since the 1990s, 4,5 at a
rate of approximately 3% per year in the recent decade.6 This rising trend of eoCRC has been noted
internationally.7–10 The causes for the rise of eoCRC,
however, remain to be elucidated.
Studies have shown that birth cohort effects are the dominating phenomenon
for the observed trend in eoCRC incidence2,7,11, suggesting etiologic roles of environmental
or behavioral risk factors that differ across birth cohorts. For example, positive
associations with eoCRC have been reported for dietary patterns (e.g., more
processed meat and less vegetables), sedentary lifestyle, smoking, obesity, and type
2 diabetes, although there have also been mixed findings.12–17
Antibiotic use has been speculated to increase the risk of CRC by causing
gut microbiome dysbiosis.18–24 Broad-spectrum antibiotics,
effective against a wide range of bacteria, contribute to intense and long-lasting
gut dysbiosis compared with narrow-spectrum antibiotics. Evidence from animal
studies supports broad-spectrum antibiotics as CRC carcinogens.25,26 In
epidemiologic studies, an association between overall (i.e., both broad and narrow
spectrum) antibiotic use and risk of average-age onset CRC has been shown in some
studies27–31. Studies also showed that the association
with CRC appeared to differ by antibiotic class32–34.
Meta-analyses pooling antibiotic class by spectrum showed that the association may
be mostly limited to broad spectrum antibiotics.27,31
To date, only a few studies have evaluated associations between antibiotic
exposure and risk of eoCRC 35–39. Two
studies examined antibiotics by class (both reported null associations for all
classes in the fully adjusted model) and spectrum in eoCRC35,39.
However, these prior findings may be difficult to interpret due to lack of
evaluation on duration of use or the relevant window of exposure. In the US, the
number of prescriptions for broad-spectrum antibiotics has increased from 0 in 1950
to 13 million in 2002. In contrast, narrow-spectrum antibiotic prescriptions
increased from 7,300 to 63,000 between these years40,41.
Despite the decline of overall antibiotic use since early 2000, 42–45 broad-spectrum antibiotic prescriptions doubled from 2000 to
2010.40 Thus, the
hypothesis that broad-spectrum antibiotic use, rather than narrow-spectrum
antibiotic use, increases the risk of eoCRC, warrants further evaluation.
A better understanding of the potential effect of antibiotic use by spectrum
could inform strategies of eoCRC control related to antibiotic use. To address
current research gaps, we evaluated the associations between broad-spectrum and
narrow-spectrum antibiotic use and risk of eoCRC in a nested case-control study
considering different antibiotic exposure windows prior to eoCRC diagnosis.

Materials and Methods

Study design, setting, and population
This study was conducted in Kaiser Permanente Southern California
(KPSC), an integrated healthcare delivery system serving over 4.6 million
members who are broadly representative of the racially/ethnically and
socioeconomically diverse residents in southern California.46,47 We used a population-based nested case-control study design
to evaluate the associations between oral broad-spectrum antibiotic use and risk
of early-onset colorectal adenocarcinoma. KPSC members who were diagnosed with
in situ or invasive CRC at ages 15–49 between
2009–2021 were eligible for this study. Cases diagnosed between
2009–2020 were identified using KPSC’s Surveillance, Epidemiology,
and End Result (SEER)-affiliated cancer registry. Cases diagnosed in 2021 were
initially identified by International Classification of Diseases (ICD)-10 codes
and manually chart reviewed to confirm the CRC diagnosis. Characteristics of the
CRC cases, including date of diagnosis, histology, stage at diagnosis, and
anatomic site were collected from KPSC’s cancer registry or chart review.
We excluded cases who: (1) had less than 2 years of KPSC membership prior to
their CRC diagnosis; (2) had another cancer diagnosis (except non-melanoma skin
cancer), human immunodeficiency virus (HIV) infection, or transplant history
prior to the CRC diagnosis; (3) had inflammatory bowel disease (IBD), familial
adenomatous polyposis (FAP), Lynch syndrome, or Peutz–Jeghers syndrome;
or (4) had histology other than adenocarcinoma (e.g., neuroendocrine or squamous
cell histology).
Each eoCRC case was individually matched up to 10 control subjects. For
each control, an index date that was the date of the matched case’s CRC
diagnosis was assigned. Matching was performed based on age (years), sex, and
years of KPSC membership (exact or +1 year in controls) prior to diagnosis/index
date, using incidence density sampling. Those who met the exclusion criteria (1)
– (3) described above were not eligible.
We further limited the primary analysis to cases and controls with at
least 15 years of KPSC membership prior to the diagnosis/index date to allow for
the evaluation of exposures from 10–14.9 years prior, a more
etiologically relevant period for CRC initiation. This study was approved by the
KPSC’s Institutional Review Board with waiver of informed consent and was
conducted in accordance with Declaration of Helsinki.

Data collection
The exposures of interest were oral broad-spectrum and narrow-spectrum
antibiotics. We extracted data on all oral antibiotic prescription fills from
KPSC’s outpatient and inpatient pharmacy databases between 2–14.9
years prior to diagnosis/index date. We limited all exposure and potential
confounder measurements to >=2 years prior to diagnosis/index date
(rather than at index date) to reduce protopathic bias, since the progression of
CRC and its symptoms could lead to antibiotic prescriptions. This approach also
reduced the chance that measurement of potential confounders was affected by
increased health system encounters for the cancer diagnostic workup in
cases.
We focused on oral antibiotic prescriptions since the effects of
antibiotics on the gut microbiome are less well characterized for non-oral
routes and likely different in magnitude across routes.48 We collected information on prescription
date, antibiotic name, route of administration, and days of supply for each
prescription. Antibiotics were initially grouped by class (e.g., cephalosporins,
macrolides, penicillins, quinolones, sulfa/trimethoprim, tetracyclines, and
others) and then by spectrum (broad vs. narrow, determined by the medical
literature, see Supplemental
Table 1).33,49–55 Broad-spectrum antibiotics are generally
active against both gram-positive and gram-negative bacteria. Antibiotic use
status was defined separately for 10–14.9 years, 5–9.9 years, and
2–4.9 years prior to diagnosis/index date. The cumulative duration of
antibiotic use by spectrum was calculated as the total number of days of use for
the spectrum (i.e., broad or narrow) during each exposure window, ignoring
overlapping prescriptions of the same drug or different drugs within the same
spectrum (i.e., the days for which prescriptions overlapped were not double
counted), following previously published method).56
The other unmatched potential confounders considered were
race/ethnicity, Medicaid enrollment (ever, never), body mass index (BMI),
diabetes, hypertension, dyslipidemia, irritable bowel syndrome, diverticulitis,
H. pylori infection, small intestinal bacterial overgrowth syndrome, smoking
(ever, never), alcohol use (ever, never), and family history of CRC. Metabolic
conditions such as diabetes, hypertension, and dyslipidemia were considered due
to their relationship with older adult CRC and potential immune
alteration.57–63 All covariates were collected
from KPSC’s electronic health records; the specific data sources and
algorithms for covariate definition are shown in Supplemental Table 2. Values of
Medicaid enrollment, irritable bowel syndrome, diverticulitis, H. Pylori
infection, and small intestinal bacterial overgrowth syndrome were defined
separately for each of the three exposure windows: 10–14.9 years,
5–9.9 years, and 2–4.9 years prior to diagnosis/index date. For
diabetes, hypertension, and dyslipidemia, the individual was considered to have
the condition if the diagnosis was made in a given window or in a prior window.
BMI, smoking, and alcohol use were not consistently collected prior to 2007; we
thus considered all values measured between 2–14.9 years prior to
diagnosis but selected the one closest to 15-year prior to diagnosis/index date
for the 10–14.9 years exposure windows. For the 5–9.9 years and
2–4.9 years exposure windows, the values within the window but closest to
10- and 5-year prior to diagnosis/index date were used, respectively.

Statistical analysis
The distributions of the demographics and clinical characteristics were
calculated and compared between the cases and their matched controls and tested
using Chi-square test or Fisher’s exact test (when >20% cell count
has expected n<5) for categorical variables. Distributions of continuous
variables were compared using the t-test. We summarized any use and cumulative
duration of antibiotic use by spectrum during each exposure window. Cumulative
duration of use was grouped into the following categories: no use, ≤30
days, 31–90 days, and >90 days. Categories with small counts
(i.e., <5 individuals) were combined with the adjacent category to form a
combined category (e.g., if “> 90 days” had <5
individuals, it would be combined with “31–90 days” to form
the “> 30 days” category).
Crude and multivariable conditional logistic regressions were used to
estimate associations between eoCRC risk and broad- or narrow-spectrum
antibiotics use in each of the three windows of exposure. To assess the
potential for confounding by antibiotic use in an earlier period, we fit a
single model with three broad-spectrum antibiotic exposure variables, one from
each exposure window. This was repeated for narrow-spectrum antibiotics.
Estimates for the association between antibiotic use and eoCRC risk were similar
when use during different exposure windows were included in a single model vs.
separate models. Multivariable models were thus fit separately for each exposure
window to preserve power. We aimed to identify a set of confounders to include
across all three exposure window models to allow for similar interpretation. We
first evaluated the associations between each potential confounder (except
race/ethnicity, which was included as a pre-specified confounder) and eoCRC in
bivariate analyses for each time window. We identified covariates with crude
p<0.10. The most extensive set of covariates with p<0.10 was for
the model for exposure 10–14.9 years prior to diagnosis/index date:
hypertension and family history of CRC. All multivariable models thus included
broad-spectrum antibiotic use, narrow-spectrum antibiotic use, race/ethnicity,
hypertension, and family history of CRC.
Both broad-spectrum and narrow-spectrum antibiotic use were modeled
first as any use, and then by the cumulative duration of use. All analyses were
repeated for colon and rectal cancer separately. A sensitivity analysis based on
those with ≥5 years of membership prior to diagnosis/index date was
conducted to assess potential selection bias associated with including only
those with ≥ 15 years of prior membership. The associations between
broad-spectrum and narrow-spectrum antibiotic exposures from 2– 4.9 years
prior to diagnosis/index date and risk of eoCRC were examined in this
sensitivity analysis and compared with those from the primary analysis. All
analyses were conducted using SAS statistical software Version 9.4; Cary, North
Carolina, USA.

Data Availability:
Anonymized data that support the findings of this study may be made
available from the corresponding author on reasonable request from qualified
researchers with documented evidence of training for human subjects
protections.

Results
A total of 2,436 individuals diagnosed with CRC between the age of
15–49 at KPSC between 2009–2021 were identified. After applying the
initial eligibility criteria, 1,334 cases with adenocarcinoma remained and were
matched to 12,858 controls. For the primary analysis, only those with >=15
years of prior enrollment were retained [295 cases and 2,794 controls, representing
3, 089 unique individuals (2 cases were chosen as controls when they were
cancer-free), Figure 1].
The mean age at diagnosis for the 295 eoCRC cases was 44.7 years (SD: 4.8
years); approximately half were male (47.5%). Cases were racially/ethnically
diverse: 31% non-Hispanic White, 17% non-Hispanic Black, 11% non-Hispanic
Asian/Pacific Islander, and 39% Hispanic (any race) (Table 1). Almost 70% of the cases had colon cancer, versus 30% with
rectum cancer. Almost half of cases were diagnosed at a regional stage (SEER summary
stage). Control subjects generally had similar demographic and clinical
characteristics distributions, except for race/ethnicity, hypertension history, and
family history of CRC (Table 1).
Between 10–14.9 years prior to diagnosis/index date, 65% of cases and
64% of controls used broad-spectrum antibiotics, while 43% and 46% of cases and
controls used narrow-spectrum antibiotics. The mean cumulative duration of use was
30 and 26 days for broad-spectrum antibiotics and 8 and 11 days for narrow-spectrum
antibiotics for cases and controls, respectively. Neither broad-spectrum (adjusted
odds ratio (aOR)=1.07 (0.82–1.40)) nor narrow-spectrum (aOR=0.85
(0.66–1.10)) antibiotic use 10–14.9 years prior to index date was
associated with eoCRC risk (Table 2). None of
the point estimates for the association between broad spectrum antibiotic use
(compared to non-use) by duration were statistically significant, although they
increased with longer duration of use (Table
2). For narrow-spectrum antibiotic use, the aORs were less than one and
were not statistically significant (Table
2).
In the analysis focusing on antibiotic exposures in 10–14.9 years
prior to diagnosis/index date and restricting to colon cancer, the OR for any
broad-spectrum antibiotic use was 1.21 (0.88–1.67). Cumulative duration of
broad-spectrum antibiotic use >90 days compared to no use was associated with
risk of developing colon cancer [aOR= 2.04 (1.09–3.85), p=0.03, Table 3]. We were unable to examine cumulative
duration of narrow-spectrum antibiotic use >90 days due to small numbers;
however, use >30 days was not associated with risk of colon cancer [adjusted
OR= 0.69 (0.36–1.29), p=0.24]. No clear associations with broad-spectrum or
narrow-spectrum antibiotic use were observed with rectal cancer (Table 3). However, we were unable to examine cumulative
use of >90 days for broad-spectrum antibiotics for rectal cancer due to small
cell count.
When we examined antibiotic use from 5–9.9 years and 2–4.9
years prior to diagnosis/index date, no clear patterns emerged for broad-spectrum
antibiotic use overall or by duration of use with overall eoCRC, colon cancer, or
rectal cancer (overall use only, due to insufficient number to model duration of
use) (Table 4). For cumulative duration of
broad-spectrum antibiotic use > 90 days, aOR was 0.67 (0.33–1.39) and
0.89 (0.34–2.33) for use 5–9.9 years and 2–4.9 years prior to
diagnosis/index date, respectively, for overall CRC, and 1. 07 (0.50–2.27)
and 1.43 (0.53–3.88, n=5 for this level of exposure), respectively, for colon
cancer (Table 4). No clear associations were
found for narrow-spectrum antibiotics, with aOR estimates ranged between
0.67–0.96 across CRC outcomes and windows of exposure (Table 4).
In the sensitivity analysis among those who had ≥5 years of
membership prior to diagnosis/index date, 947 cases and 9,080 controls were included
(Supplemental Table 3).
The adjusted OR for broad-spectrum antibiotic any use, ≤30 days, 31–90
days, and >90 days during the 2–4.9 years prior to diagnosis/index
date were 0.94 (0.82, 1.08), 0.94 (0.81, 1.09), 0.94 (0.70, 1.27) and 0.91 (0.54,
1.52), respectively, for overall eoCRC. These ORs for narrow-spectrum antibiotic any
use, <=30 days and >30 days were 0.95 (0.81, 1.12), 0.95 (0.80, 1.12)
and 1.07 (0.61, 1.89), respectively (Supplemental Table 4).

Discussion
In this population-based nested case-control study, we investigated the
association between antibiotic use and eoCRC diagnosis by antibiotic spectrum and
exposure window. We observed a positive association between early-onset colon
adenocarcinoma and broad-spectrum antibiotic use for > 90 days in cumulative
duration from 10–14.9 years prior to diagnosis. In this same etiologic
window, we did not observe an association between broad-spectrum antibiotic use and
rectal cancer. No clear patterns were present for broad-spectrum antibiotic use
between 2–4.9 or 5–9.9 years prior to diagnosis/index for all three
outcomes evaluated (overall eoCRC, colon cancer, and rectal cancer). While the odds
ratios were generally lower than one for narrow-spectrum antibiotic use, no
statistically significant associations were found for narrow-spectrum antibiotic use
in any analysis.
Several systematic review and meta-analyses have shown a positive
association between overall antibiotic use and risk of average-onset CRC27–31. Of these, one meta-analysis of antibiotic classes by
spectrum reported a pooled relative estimate of 1.70 in effect size (95% CI:
1.26–2.30) for broad-spectrum antibiotics, in contrast to the weaker,
non-significant association for narrow-spectrum antibiotics [pooled effect estimate
=1.11 (0.93–1.32)].31
However, in another meta-analysis pooling estimates by antibiotic classes, not all
broad-spectrum classes examined had a positive, significant pooled
estimate.27 It should be
noted that the pooled estimates represented summary estimates of antibiotic
exposures from varying exposure windows. For most studies, the overall estimates
were derived from antibiotics exposures that included those within 10 years prior to
diagnosis. It is possible that the differential associations seen for different
classes of broad-spectrum antibiotics may be influenced by trends of common
antibiotic classes in use and differences in etiologic windows. Standardizing by
etiologic windows of exposure may be a useful exercise for future meta-analyses that
evaluate CRC etiology.
To date, only a few studies have evaluated the association between
antibiotic use and risk of eoCRC. Nguyen et al. included 2,557 eoCRC cases from
Sweden and collected antibiotic use data from 6 months to 10 years prior to
diagnosis.35
Broad-spectrum antibiotics use [aOR=1.13 (1.02–1.26)], but not
narrow-spectrum antibiotic use [aOR=1.01 (0.92–1.11)], was associated with
risk of eoCRC. However, no dose-response relationship was observed, and no clear
association was found by CRC location. In the case-control study by McDowell at al
(eoCRC n=445 from Scotland) a positive association between antibiotic use (from
>1 year prior to diagnosis) and eoCRC was reported. However, no dose-response
was observed, and they did not evaluate broad vs. narrow spectrum antibiotic
exposure.36 Jiang et al.
reported a positive association between long-term antibiotic use during early life
(self-report) and risk of eoCRC [n=165, OR=1.48 (1.01–2.17)] using UK biobank
data.38 They also reported
potential gene-environment interaction for this antibiotic exposure. However, data
on antibiotic spectrum were not available. Ben-Aharon et al screened across
medication classes for their associations with eoCRC risk (eoCRC n=941, from
Israel), and did not find an association for antibiotics, yet they did not evaluate
antibiotic classes, spectrum, or dose response.37 A study conducted in Kaiser Permanente Northern California
(a health care system similar to KPSC) reported no clear association with antibiotic
use overall, by class or spectrum from >1 year prior to diagnosis.39 In a subset analysis, they
examined antibiotic use up to ≥8 years prior to diagnosis and found no clear
association for any use vs. no use. In contrast, we designed our study to test the
hypothesis of broad- vs. narrow-spectrum antibiotic use from different windows of
exposure up to 15 years prior to diagnosis. Our data suggest that the window to
assess antibiotic exposure in relation to the timing of diagnosis may be important,
and the need for additional studies to evaluate whether duration of broad-spectrum
antibiotic use is associated with eoCRC risk 10 or more years later.
With respect to the relevant exposure window, Zhang et al found a positive
association between the risk of average-onset CRC and antibiotic use from more than
10 years prior to diagnosis, but not from 1–10 years prior to
diagnosis.33 In the study
by Armstrong et al., the strongest association between antibiotic exposure and risk
of average-onset CRC was for exposures up to 15 years prior to diagnosis, compared
to exposure within 5 years prior to diagnosis.64 A prior cohort study examining colorectal adenoma reported
a positive relationship for self-reported long-term (> 2 months) antibiotic
exposure from up to 20–40 years prior, and no association for antibiotic
exposures within the 4 years prior to diagnosis.65 A systematic review also noted similar observations - with
stronger associations seen for less recent than more recent exposure in relation to
diagnosis.28 Taken
together, these data suggest that prolonged antibiotic exposure likely play a role
in CRC initiation, with a lesser extent, if any, for disease progression. These data
also suggest that future studies should carefully and jointly consider the exposure
data collection window, antibiotic spectrum, and duration of exposure.
Potential difference in findings between colon and rectal cancer have been
reported by several prior studies in average-onset CRC,27,32,33 and one study for eoCRC.36 Lu and colleagues reported a small
non-significant inverse association between antibiotic exposure and rectal cancer
[OR=0.91 (0.80–1.04)], and a more pronounced association in women [OR=0.86
(0.80–0.92)].66 It
has been speculated that colon’s proximity to small intestine and the
presence of colon biofilms (which may increase susceptibility to antibiotic-induced
dysbiosis) may explain the differential risk patterns observed67. In this study we were unable to examine
> 90 days of broad-spectrum antibiotic use for rectal adenocarcinoma due to
small cell sizes. Future research should continue to shed light on the role of
antibiotic exposure on CRC by anatomical location.
Broad-spectrum antibiotics are effective against both gram-positive and
gram-negative bacteria and multiple bacteria groups, thus contributing most to gut
dysbiosis compared with narrow-spectrum antibiotics.68–70 Characterized by the loss of beneficial bacteria and enrichment
with pathogenic bacteria, gut dysbiosis has been strongly implicated in CRC
carcinogenesis.71 Although
the carcinogenic mechanisms remain to be fully elucidated, hypothesized mechanisms
include bacterial direct production of carcinogenic and genotoxic compounds;
bacterial-induced chronic inflammation; promotion of DNA damage and loss of
apoptosis; and through epigenetic changes favoring cancer development.72,73 A recent study found that mutational signatures associated
with colibactin (a genotoxin produced by pks+ E. coli) was enriched in
eoCRC.74,75 It is possible that the increased use of
broad-spectrum antibiotics may play a role in favoring the selection of E. coli in
human guts76, thus promoting the
increase of eoCRC. Human studies have demonstrated almost complete eradication on
microbial composition after exposure to broad-spectrum antibiotics, 69,77 and the recovered microbiome often suffers loss in
beneficial commensals.69,77,78 In
animal studies, long-term administration of broad-spectrum antibiotics induces CRC
formation.25,26 The potential biological mechanisms and
animal data support the need to further evaluate antibiotic exposure by antibiotic
spectrum.
Several potential limitations should be considered when interpreting our
results. First, to assess antibiotic exposure in a more etiologically relevant
window, only those with a long KPSC membership history (i.e., ≥ 15 years)
were included, excluding a large proportion of the eoCRC cases. That said, we do not
expect important bias from this exclusion as our cases and controls were matched on
membership length. In the sensitivity analysis including cases and matched controls
with at least 5 years of prior membership showed similar null findings for
antibiotic exposures from 2–4.9 years prior to diagnosis, although OR
estimates were closer to null for narrow-spectrum antibiotics in this sensitivity
analysis. Another implication of the long membership requirement was that the cases
included were somewhat older than the overall eoCRC cases. However, there were not
enough cases diagnosed at the younger age (i.e., under age 40) to allow proper
evaluation of effect modification by age at diagnosis. Second, there remains the
possibility of residual confounding due to potential misclassification of BMI and
smoking in earlier study years, and of alcohol use due to lack of standard
definition, and unmeasured confounders such as physical activity, sedentary
behaviors, and diet. Third, the number of individuals with cumulative exposure of
>90 days for both broad- and narrow-spectrum antibiotic use was limited,
calling for caution when interpreting these results. The analysis of rectal cancer
was also based on a relatively small number of cases (n=90). Fourth, we acknowledge
multiple comparisons as a limitation as we examined exposures from three windows of
exposure and different anatomical locations. Fifth, due to limited sample size, we
did not examine antibiotic spectrum and class jointly, although the potential
effects of broad-spectrum antibiotics may vary by class. We also did not examine the
potential relationship with cumulative dose. Finally, we did not evaluate other
routes of antibiotic use besides oral use. Our electronic health record data on
intravenous (IV) antibiotic use was incomplete during the study period. That said,
IV antibiotic use is expected to be relatively rare in this young study population
and is unlikely an important risk factor for driving the rising trend of eoCRC. It
should also be noted that our definition of activity spectrum was primarily based on
current understanding of the effective bacterial targets, which were often
pathogens. Emerging evidence on the impact of different antibiotics on gut
microbiota may lead to better classification for the purpose of understanding risk
factors for eoCRC.
Our study has several important strengths. KPSC’s stable membership
allowed us to study antibiotic exposures from different exposure windows up to 15
years prior to diagnosis, providing robust U.S. data while prior literature weighted
heavily from countries with national health databases. KPSC’s comprehensive
pharmacy database permitted detailed capture of antibiotic types and days of
prescription. Further, we were able to ascertain important clinical conditions such
as H. pylori infection to address confounding by indication, and we evaluated
exposures from at least 2 years prior to diagnosis to minimize protopathic bias.
A potential positive association between >90 days cumulative
broad-spectrum antibiotic use and early-onset colon cancer suggested by the data
calls for further evaluation in larger studies. Well-designed epidemiologic studies
leveraging large, long-term population-based health databases will be critical to
help investigate these associations, specifically, studies that focus on antibiotic
spectrum and duration of use during specific exposure windows.

Supplementary Material
1234