Regular use of paracetamol and risk of liver cancer: a prospective cohort study

The United Kingdom Biobank is a large-scale, prospective, population-based cohort of over 500,000 individuals aged 37–73 years who were recruited from 21 assessment centers across the U.K. in 2006–2010. All eligible participants were invited to complete touchscreen questionnaires, face-to-face interviews, physical measurements, and biological sample collection. Detailed information about the project is available on the website (https://​www.​ukbiobank.​ac.​uk/​) and in previous studies [11]. The UK Biobank was approved by the North West Multi-Center Research Ethics Committee. All participants provided written informed consent prior to data collection. In this study, we excluded 36,865 participants with a diagnosis of cancer and 1,301 participants who had missing data on other covariates. The final analyses included 464,244 participants.

At baseline, regular use of paracetamol was first evaluated by participants using a touchscreen questionnaire and subsequently confirmed by a UK-biobank trained staff. “Regular use” was defined as taking the medication in most days of the week for the last 4 weeks. Information regarding the doses and duration of paracetamol use was not collected.

Incident liver cancer cases within the UK Biobank cohort were identified by ICD-10 codes C22. This information was recorded from cancer and death registries from the Health and Social Care Information Centre (in England and Wales) and the National Health Service Central Register (in Scotland). The diagnosis of liver cancer was confirmed by medical records, pathology reports, imaging results, and death certificates. Details of the methods can be found on the UK Biobank website.

Covariates were determined from a touchscreen questionnaire and verbal interview at baseline. These included sociodemographic factors (age, sex, and race) and lifestyle factors (smoking, alcohol consumption, sleep time, and diet habits). The index of multiple deprivations based on the postcode of residence was determined as a composite measure of socioeconomic status. Physical activity was estimated by the validated Short International Physical Activity Questionnaire (IPAQ) for all individuals. Comorbidities (hypertension, diabetes, hyperlipidemia, viral hepatitis, cirrhosis) and medication use (multivitamins, mineral supplements, aspirin, ibuprofen, NSAIDs, PPIs, histamine-2 receptor antagonists, antihypertensive drugs, antidiabetic drugs, and statins) were assessed based on self-reported medical history. Body mass index (BMI) was calculated by dividing weight by the square of height (kg/m2).

Person-years were calculated from the date of return of the baseline questionnaire to the date of first diagnosis of liver cancer, death, or the end of follow-up (31 December 2020), whichever came first. Cox proportional hazards models with age as the timescale were fitted to calculate the hazard ratios (HRs) with 95% confidence intervals (CIs).

We employed an overlap propensity score weighting approach to address potential confounding. First, we used multivariate logistic regression model conditioned on baseline covariates, including age, sex, centers, race, socioeconomic status (index of multiple deprivation), smoking status, alcohol consumption, physical activity, fruit and vegetable intake, meat intake, sleep time, BMI, concomitant comorbidities (hypertension, diabetes, hyperlipidemia, viral hepatitis, cirrhosis), current medication (multivitamins, mineral supplements, aspirin, ibuprofen, NSAIDs, PPIs, histamine-2 receptor antagonists, antihypertensive drugs, antidiabetic drugs, and statins), overall health rating, long-standing illness, and family history of cancers, to calculate the propensity score for paracetamol use. The overlap weight based on the propensity scores was then applied to establish a pseudo population in which the measured confounders were balanced between paracetamol users andnonusers. Standardized mean differences (SMDs) were calculated before and after weighting to assess covariate balance, with SMD less than 0.1 considered negligible [12]. A weighted Kaplan–Meier curve was generated to characterize the cumulative incidence of liver cancer over time. Weighted Cox regression models were used to estimate marginal HRs with 95% CIs for the effect of paracetamol use on liver cancer risk. Schoenfeld’s tests were employed to check the proportional hazard assumption and no violation was detected. To present the association easily, we calculated the number needed to harm (NNH) with the method described by Altman and Andersen [13].

To further investigate potential effect modifiers, we conducted subgroup analyses stratified by sex, age, obesity, smoking and drinking status. We performed several sensitivity analyses to test the robustness of our findings. First, we performed a lagged analysis of the exposure for 2 years to minimize the potential for protopathic bias. Second, we excluded participants with viral hepatitis and cirrhosis to control the potential impact of health conditions. Third, we employed an alternative inverse probability treatment weighting (IPTW) approach to further mitigate the impact of confounding variables. The IPTW method aimed to achieve a balanced distribution of measured confounders between paracetamol users and nonusers. The propensity scores for paracetamol use were derived from a multivariable logistic regression model. We evaluated the balance of covariates between the two groups by computing the standardized mean differences (SMDs) before and after weighting. Covariates with SMDs less than 0.1 were considered negligibly unbalanced. Weighted Cox regression models were used to estimate the HRs and 95% CIs as described by Austin et al. [14]. Last, we calculated the E-value to estimate the potential role of unmeasured confounders, which represents the minimum strengths of association between an unmeasured confounder and exposure or outcome that can fully explain away a specific treatment–outcome association, conditional on the measured covariates [15]. All analyses were conducted using R software (version 4.1.0, R Foundation for Statistical Computing, Vienna, Austria).