The association of antimicrobial use with the increasing development of antimicrobial resistance has been previously described and is now widely considered to be a serious global health problem(1). Use of antibiotics can upset the human gut microbiome, resulting in acute disease. However, treatment of the pathogenic organism with antibiotics can result in a cure. Replacement of the disrupted microbiome with ‘normal’, non-antibiotic treated gut flora in the form of fecal matter transplants can also yield excellent therapeutic results, as in the case of C. difficile diarrhea(2).
However, the role of the microbiome as an etiologic single entity factor in human disease is a relatively uncharted territory. The idea that a dynamic and inter-related community of diverse micro-organisms living in discrete body compartments, such as the gut and respiratory tract, might be involved in the pathogenesis of disease has been investigated in only a few conditions including:
- Inflammatory bowel disease(3)
- Colorectal cancer(4)
- Allergic and immune diseases including asthma(5)
- Obesity (6)
- Vascular disease (7).
Much of this work has been at the molecular and cell signaling pathways level; however, there is sparse epidemiological evidence that long term disturbance of the microbiome might contribute to the development of chronic disease. Here, we study the long-term use of antibiotics that may alter the microbiome in relation to chronic disease, looking for epidemiological evidence of the association between antimicrobial use and the development of asthma, rheumatoid arthritis, inflammatory bowel disease, colorectal cancer, and dementia (including vascular dementia).
Subjects and setting
This retrospective observational study data set was drawn from the RCGP RSC, which hosts data as a pseudonymized [virtually anonymized – no names are held but demographic and medical record data are; identification is virtually impossible; data is encrypted and stored on dedicated secure servers at the University of Surrey. The secure network at University of Surrey meets the NHS information security standard for holding these data as set out in the Information Governance toolkit] dataset from a nationally representative sample of 644,273 people registered with the network practices of all ages (range approximately 0 – 100) (8). The database has been involved in surveillance of influenza and respiratory disease for over 50 years. Over this period, practices have had feedback about their data quality, in particular the differentiation of new (incident) cases from follow-up consultations. Data quality is as good as any other existing dataset held for routine primary care(9).
UK general practice is suitable for this type of study because it has a registration-based system with patients registered with a single practice. Practices have been computerized since the late 1990s, with pay-for-performance introduced in 2004 for chronic disease management. Key data are coded 10, which include diagnoses, therapy, test results, and other key data. Ethics approval was not required for this study, as determined by the Medical Research Council Heath Research Authority decision tool.
Antibiotic codes were searched using the UK’s most ubiquitous electronic medical record for primary care, Egton Medical Information Systems (EMIS) and Read v2 drug codes, to capture prescribing using the different Electronic Medical Record (EMR) systems. Suitable diagnoses and known risk factors were defined ontologically.
Association of chronic antibiotic prescribing with disease was tested by interrogating the dataset for new cases of asthma, colorectal carcinoma, rheumatoid arthritis, dementia, inflammatory bowel disease, and known risk factors for these diseases. New cases were defined as records in which the first recorded diagnosis in the EMR appeared in 2015, represented by the recording of clinician assigned diagnostic codes in the EMR in that year. Antibiotic load was defined as the sum of antibiotic prescriptions issued in the preceding ten years. Dates were chosen to ensure that 10 years of complete data were available.
Risk factors, e.g. smoking, diabetes, hypertension, and relevant family history were identified through literature searches and ontologically defined. Records were searched for the occurrence of risk factors at any point in time within the subject record.
New cases were combined with demographic data, known risk factors for the diseases in question, and the number of antibiotic prescriptions in each of the preceding ten years to create the diseases-specific datasets and total antibiotic load in the entire period (sum of each of the preceding ten years’ prescriptions). The independent risk factors for the diseases are shown in Table 1. The selection of subjects for regression model data frames is shown schematically (Figure 1).
Multiple logistic regression analysis was undertaken using R studio version 1.0.153 to control for demographic variables and known risk factors.
The logistic regression model yielded an estimate of the odds ratio (OR) of being diagnosed with one of these diseases (defined as first diagnosis in the record in 2015) regressed against antibiotic prescribing load (sum of antibiotic prescriptions in the preceding decade), demographic variables, and other independent variables, i.e. known risk factors for the disease in question.
Logistic regression models were run against data frames that maximized the number of records; for example, if alcohol was deemed to be a risk factor that needed to be controlled for, records that had no record of alcohol intake were removed; otherwise the full dataset was used where each subject had the risk factor satisfactorily recorded. Records with no ethnicity, smoking, or BMI data were removed entirely. The basic logistic regression model fitted to the data is:
Disease First Recorded in 2015 ~ Total Antibiotic Load 2004-2014 + Ethnicity + Gender + Age + Smoking Status + IMD Score + BMI + Specific Disease Risk Factors
Log [(probability disease diagnosed in 2015)/(1-probability disease diagnosed in 2015)] = ß0 + ß1AntibioticLoad + ß2Age + ß3Ethnicity + ……… + ßXRisk FactorX
The baseline dataset comprised 644,273 records. This was reduced to 444,115 when alcohol intake was included in the models (records with no alcohol intake coding were removed) and 530,695 with only smoking and BMI cleaned data (records with no BMI or smoking data recorded removed). Sample characteristics are shown in Table 2. There are some previously described differences between the RCGP RSC database and national population characteristics, e.g., the RCGP RSC population is slightly less obese and less deprived (as defined by the Index of Multiple Deprivation score, widely used in the UK), but disease prevalence is mostly identical to the national population (8). The total antibiotic prescription load for the decade is summarized in Table 3. Prescribing by year did not vary significantly from year to year.
The ORs of developing diseases after adjusting for known risk factors are shown in Table 4. The OR is the increase in odds per antibiotic prescription over the preceding decade, adjusted for age, gender, IMD decile and other known risk factors for the diseases. Table 3 shows sample characteristics compared to the sampling frame and national population.
The complete OR of disease development for each risk factor is shown in Tables 5-9. ORs in bold indicate significant risk factors at the 5% level.
Reference levels are: Asian, smoker, female, IMD Decile 1 (poorest), alcohol moderate (where applicable to model – inflammatory bowel disease, dementia and bowel cancer models).
This study demonstrates a statistically significant, dose-related, positive association between the odds of being diagnosed with asthma, rheumatoid arthritis, and inflammatory bowel disease and the total amount of antibiotics prescribed in the preceding decade in a representative sample of UK primary care patients, across all age ranges. The increase in odds of diagnosis is per antibiotic prescription, so an apparently small increase in the odds ratio becomes much larger in subjects who have received more prescriptions. For example, the odds of developing asthma, at 1.004, increase to 1.40, i.e., by 40% over the 10-year time period if 100 antibiotic prescriptions in the preceding decade were issued (equating to 10 prescriptions per year). No such association was seen for colorectal cancer or dementia. On a population basis, this implies that if there is indeed a causal relationship, that antibiotic prescribing may be a significant driver of chronic disease.
A possible common link is modulation of the microbiome and its long-term disruption by repeated administration of antibiotics. Relevant mechanisms might include the overgrowth of pathogenic organisms (as in the case of antibiotic associated pseudomembranous colitis), the disturbance of particular metabolic pathways to produce pathogenic molecules and the decimation of protective organisms. If proven, it may be that a proportion of total cases of these diseases is iatrogenic. The implications add weight to efforts to reduce antibiotic use and may be a useful public health message for policy makers to inform the public and perhaps help change patient expectations during physician encounters. More than that, if there is a true causal association between antibiotic prescribing and chronic disease, the finding is relevant to prescribers, researchers considering the mechanisms of disease causation, regulators governing the issuance of antibiotics, the government and the pharmaceutical industry undertaking post-marketing surveillance of drug safety.
The demonstrated association is a statistically significant dose-related increase in the OR of being diagnosed with asthma, rheumatoid arthritis, or inflammatory bowel disease across all age ranges in a sample of subjects drawn from patients registered with English general practices. Potential mechanisms include chronic disruption of the microbiome. This finding has implications for practitioners who prescribe antibiotics, the pharmaceutical industry, policy makers, and researchers involved in studying chronic disease mechanisms. In summary, this study demonstrates a statistically significant positive association between the long-term use of antibiotics and the diagnosis of asthma, inflammatory bowel disease and rheumatoid arthritis which should be explored further by both epidemiological and basic sciences research.
- Patient de-registration during the period would bias the results towards the null as these patients would not have been coded as cases but might have had prescriptions. This is unlikely to be a significant effect, given the small number of de-registrations during the study period. Additionally, the results of three of the models are statistically significant despite this bias and may be greater in magnitude if it were not present.
- It has been assumed that all the prescriptions issued were actually ingested as issued. If they were not actually taken, it would be hard to ascribe the effects described to the antibiotics.
- The study grouped all antibiotics into one category. Given that antibiotics are in fact a very heterogeneous group of drugs, it may be that this biased the results towards the null. One might intuitively expect broad spectrum antibiotics to have a greater effect on changing the microbiome, but equally, narrow spectrum drugs might affect a particularly important microbe. It would be interesting to undertake further studies looking at individual drugs or groups of drugs by spectrum of activity.
- The removal of some records might have led to some bias in the study. The characteristics of the removed data were examined in the case of those records which had no alcohol consumption recorded and were found to be a much younger group of patients, which might explain why no consumption had been noted by primary care practitioners. This subset would have been most unlikely to have developed dementia or bowel cancer, thus biasing the result away from the null. However, given the results for which these records were removed were non-significant in any case, it seems likely that inclusion of the records would not have affected the estimated ORs. As alcohol is not deemed to be a risk factor for asthma or rheumatoid arthritis, this would not have biased the results for these conditions.
We would like to thank all practices who have agreed to be part of the RCGP RSC and allowed us to extract and use health data for surveillance and research, and other members of the Clinical Informatics and Health Outcomes Research Group at University of Surrey, particularly Professor Simon Jones for his statistical expertise and input on methods and Jeremy Van Vlymen for his expertise in R coding.
Conflict of interests: None to be declared.
Funding and support: None.