Multi-morbidity and its association with common cancer diagnoses: a UK Biobank prospective study

UKB is a large, prospective cohort study in which ~ 500,000 volunteers underwent a comprehensive baseline assessment between 2006–2010, providing in-depth data on socio-demographics, lifestyle and health, and a range of physical measures. Participants consented to linkage to their electronic health records for longitudinal follow-up of health outcomes. Full details of the UKB cohort are given elsewhere [19].

Linkages have been undertaken to national cancer (data available from 1979 onwards) and death (2006 onwards) registries, hospital inpatient data (1997 onwards) and, for ~ 45% of the cohort, primary-care records (1938 – 2017).

UKB received ethics approval from the National Information Governance Board for Health and Social Care and the National Health Service North West Multicentre Research Ethics Committee (21/NW/0157).

In the current analysis, only UKB participants with no evidence of pre-existing cancer were included in the study.

The CMS [18] is a weighted multi-morbidity score, which is formed of 37 different conditions selected for their substantial impact on patients, based on expert opinion [18]. For the purposes of this study, four of these conditions were excluded from the assessment of multi-morbidity, including cancer (since this was the outcome of interest), vision/ hearing loss and painful conditions (as not well defined in UKB), and learning difficulties (owing to its low prevalence due to the consenting procedures for UKB). The remaining 33 conditions were used to determine the CMS in this study.

Multi-morbidity at recruitment was defined in multiple ways, as follows: i) ≥ 2 diseases, ii) weighted CMS (using published weights [18]) categorised into fourths (with those with no diseases as a separate category), and iii) a categorical variable representing number of conditions (0, 1, 2, 3 and ≥ 4). This latter measure of disease count was used as the primary exposure of interest owing to its ease of interpretation.

Code lists for disease definitions are available at https://​doi.​org/​10.​5281/​zenodo.​7334984.

The cancer registry data was used to identify cases during the follow-up period. Participants with a cancer diagnosis prior to recruitment (excluding non-melanoma skin cancer) were considered prevalent cases and excluded from the analysis. The first incident cancer diagnosis (excluding non-melanoma skin cancer) was identified for remaining participants. The four most commonly diagnosed cancers in UKB were selected as outcomes of interest because of their high public health impact and large numbers of incident cases. Prostate cancer was defined as ICD-10 code C61, breast cancer as C50, lung cancer as C33-34 and colorectal cancer as C18-20.

Date of birth was estimated from month and year of birth. Townsend deprivation score (an area-based score denoting socio-economic deprivation of a participant’s residence at recruitment) was categorised into fifths. Region was defined as London; South East; South West; East Midlands; West Midlands; North East; North West; Yorkshire and the Humber; Scotland East; Scotland West; Wales. Ethnicity was characterised as white and non-white.

Body mass index (BMI) was categorised as < 25; 25–30; ≥ 30 kg/m². Smoking was categorised as never; former; < 10 cigarettes daily; ≥ 10 cigarettes daily; current smoker with unknown number of cigarettes daily. Alcohol consumption was categorised as never; special occasions only; 1–3 times per month; 1–2 times per week; 3–4 times per week; daily or almost daily. Menopausal hormonal therapy (MHT) use was categorised as never or ever users (women only).

Self-reported screening behaviour was categorised as never or ever attended breast or bowel cancer screening and prostate specific antigen (PSA) testing. To further assess the extent to which overall health care utilisation might influence the association between multi-morbidity and cancer diagnosis, the following data were extracted from the primary-care records at recruitment: rapid referral for cancer diagnostics (yes or no), number of consultations (i.e. number of times a participant had attended their primary-care clinician; categorised as fifths), number of spirometry measurements, number of blood pressure measurements, and number of PSA tests (each categorised as 1, 2, 3, 4, ≥ 5).

Individuals contributed person-years to the analysis from their date of recruitment (2006–2010) until the earliest of: date of first cancer diagnosis (excluding non-melanoma skin cancer), loss to follow up (e.g. withdrawal from study or leaving the U.K.), death, or end of follow up. End of follow up was defined as the date for which cancer registry data were considered complete (England and Wales 29^th February 2020; Scotland 31^st January 2021).

Distribution of multi-morbidity measures according to various characteristics were calculated and summarised as median and interquartile range (IQR), mean and standard deviation (SD) or proportions.

Cox proportional hazards models, with attained age as the underlying time variable, were used to investigate the association between the multi-morbidity metrics at recruitment and subsequent cancer diagnosis during the follow-up period. Initial investigations compared these associations across follow-up intervals (< 1, 1–5, ≥ 5 years) and heterogeneity was assessed using likelihood ratio (LR) tests. All subsequent analyses focused on associations of multi-morbidity with cancer risk in the period 1 or more years after the recording of morbidity data (i.e. excluding the first year of follow-up) to minimise the impact of reverse causation.

Initial models were adjusted for year of birth, region, socio-economic status and ethnicity. Proportionality was checked by examining Schoenfeld’s residuals. To investigate the possibility of residual confounding (i.e. to determine if any association might be partly explained by the presence of shared risk factors), the models were further adjusted for BMI, alcohol consumption, smoking and MHT use. The change in LRχ²statistics associated with multi-morbidity were estimated before and after adjustment for these confounders [20]. For the purposes of this study, a reduction in the LRχ² of more than two-thirds after adjustment for all potential confounders was taken to indicate that the adjusted association could plausibly be due to residual confounding (caused by imperfect adjustment for the confounder(s) under study). Given the particularly strong association between smoking and lung cancer and other conditions included in the CMS, analyses of multi-morbidity with lung cancer was also restricted to never smokers to further investigate the possibility of residual confounding. To assess if ascertainment bias might have influenced the association, breast, colorectal, and prostate cancer models were adjusted for self-reported routine screening behaviour.

In participants for whom primary care information were available the mean number of consultations, spirometry measurements, blood pressure measurements and PSA tests (men only) were calculated stratified by the number of diseases, in those without a cancer diagnosis. The association of multi-morbidity with cancer diagnosis among those with primary-care data was then further adjusted for the number of consultations (as a proxy of health care utilisation) – and for prostate cancer, the number of PSA tests – to determine the extent to which healthcare usage might explain the associations.

Where significant associations of multi-morbidity with cancer diagnosis were identified, associations between individual diseases within the CMS and cancer risk were investigated using the same methodology as for the multi-morbidity metrics. Multiple testing was adjusted for using the false discovery rate [21].

As it is common for a number of conditions to be diagnosed on the cancer diagnostic pathway [12], and it is also possible that patients with more diseases are more likely to be investigated for cancer, it is important to consider the effects of reverse causation on these findings (i.e. that the disease is more likely to be diagnosed because of tests being undertaken as part of the cancer diagnosis, or that an undiagnosed cancer causes the disease, which is diagnosed prior to the cancer). Indeed, there was a notably greater increase in the risk of 3 of the 4 cancers considered here in the first year of follow up, compared with later periods, suggesting this is likely. However, there were no obvious changes in cancer risk across subsequent follow-up periods, indicating that any effect of reverse causation was largely confined to the first year of follow-up.

Given the known contribution of certain modifiable lifestyle factors, such as obesity and smoking, to many of the diseases included in the CMS and some of the cancer outcomes, it is important to investigate if shared risk factors explain any observed associations. In this study, adjustment for potential shared risk factors suggests that residual confounding from these factors are unlikely to explain the associations. Further, the association between multi-morbidity and lung cancer was unaffected when limited to self-reported never-smokers, suggesting that this association is not fully explained by smoking. Although it is possible that passive smoking might have contributed to the increased risk, it is highly unlikely to explain all of the increased risk, given the magnitude of the association.

The association with lung cancer was also evident when using primary-care data to ascertain multi-morbidity, despite the differences in the distributions of diseases between the two data sources, and there was no single condition that appeared to underlie this association. Furthermore, additional adjustment for metrics of health care utilisation (to account for the possibility that multi-morbid participants may be more likely to be investigated for symptoms simply due to attending a health care provider more frequently) did not materially affect the association with lung cancer, indicating that it is unlikely to be explained by ascertainment bias. Further research on stage of lung cancer would be useful to confirm this is the case.

Whilst no association was identified between multi-morbidity and prostate cancer using the self-reported health data, an inverse association was observed when using the primary-care data. Although this could be due to ascertainment bias (particularly given the higher prevalence of PSA testing in the UKB cohort compared with the general UK population [27, 28]), the difference in mean number of PSA tests was less than 1 between those with no diseases and those with ≥ 4 diseases, so PSA testing is unlikely to fully explain the apparent reduction in prostate cancer risk. Further, the inclusion of PSA testing as a covariate did not change the risk estimate materially. Instead, it is possible that prostate cancer might be diagnosed earlier in those with fewer diseases (as those with multi-morbidity may delay getting symptoms investigated, or the symptoms may be attributed to an already diagnosed disease, such as prostate disorder), leading to an apparent lower risk in prostate cancer in those with more diseases. However, data on cancer stage is required to investigate this further.

Previous research into multi-morbidity as a risk factor for cancer has been limited, with most research focussing on the association with survival following cancer diagnoses [29, 30], the prevalence and/or risk of multi-morbidity in cancer patients [2, 31], or the impact of multi-morbidity on living with cancer (including treatment options) [9, 32]. A cross sectional study in the UK found that almost half of cancer patients have at least 2 morbidities (from a list of 11) at cancer diagnosis [4], which was particularly apparent for those with gallbladder, liver, lung, myeloma and renal cancer when compared to those with colorectal cancer [4]. A US cross-sectional study identified an association between multi-morbidity (defined as ≥ 2 diseases from a list of 6) and all site, bladder and cervical cancer, but no association with colon, lung, breast or prostate cancer [3]. However, different methodology (including diseases included and a lack of data on when diseases were diagnosed in relation to cancer) may explain the differences in results seen in the current study. A recent prospective study, also using UKB data, identified no association between multi-morbidity (defined as ≥ 2 diseases from a list of 43) and colorectal cancer risk [16]. Another recent UKB study identified no association between multi-morbidity, defined as disease count or as disease clusters, and breast cancer risk [17]. Both of these studies are consistent with our findings despite differences in cohort selection and the diseases included in the multi-morbidity definition. As the results seen in our study appear to be robust to reverse causation, residual confounding and ascertainment bias, it is possible there is an aetiological role for multi-morbidity in lung cancer risk. Both asthma [33] and COPD [34] have previously been identified as risk factors for lung cancer, and both of these conditions (in addition to several others) were found to be associated with an increased risk in our study, and so it is possible that these diseases are important contributors to the association. It is also possible that multi-morbidity results in a general increase in systemic inflammation [15], which could predispose to further disease.

This study has many strengths, including the large sample, prospective analyses and use of a validated score. As far as we are aware, this is the first study to use a multi-morbidity score (the CMS [18]) to define multi-morbidity for the investigation of cancer risk. This score was originally developed using UK primary-care data, and the disease list was determined using a combination of prevalence, impact on health care usage and impact on the patient [35]. Additionally, results were compared to those generated from primary-care records for a subset of the cohort, and despite the different distribution of some of the health conditions, the association for lung cancer remained consistent. However, there are some limitations to the study. The main analyses were undertaken using self-reported data (as not all participants had primary-care data available) but whilst it is possible that individuals may not always accurately report their morbidity status, this is of less of a concern for symptomatic diseases or those that result in a lifestyle change [36], as can be seen when comparing the prevalence of diseases from the self-reported and primary-care data in this study. Secondary-care data are available in UKB, but was not used here as some of the diseases included in the CMS are not well captured in hospital in-patient records, and so could be under ascertained. The use of self-reported data meant that changes in multi-morbidity over time could not be taken into account in the analysis. Further, the use of composite scores or disease counts may not be that informative in terms of understanding which diseases or disease patterns and associations are underling any increased risk identified. Misclassification of smoking status cannot be ruled out (as it was self-reported), however adjustments for smoking did not substantially change the point estimate for lung cancer diagnosis risk, and the association was still evident in those who reported never smoking Finally, owing to the voluntary nature of the study, UKB participants are not representative of the general population [28, 37, 38]. As such, it is not appropriate to generalise prevalence estimates of the individual diseases reported here to the general population, although there is no reason to expect that the observed association of multi-morbidity with lung cancer would not be generalizable to the wider population. As an observational analysis, causality cannot be determined from these results, and there is the possibility that unmeasured confounders, or other biases, could have contributed to the associations identified.