Multi-Morbidity and Risk of Breast Cancer among Women in the UK Biobank Cohort.

The UK Biobank is a prospective population-based cohort that recruited 273,375 women, aged 40 to 69 years, from March 2006 to July 2010 [ 28 ]. Individuals were invited to participate on a voluntary basis and provided electronic informed consent for data provision and linkage. The baseline data assessment included self-reported data on personal and family medical history, lifestyle, hormone-related factors, and sociodemographic characteristics. Additional anthropometric measurements were performed. The cohort additionally retrieved individual information from the national cancer and death registries. We excluded women with any cancer diagnosis prior to baseline except non-melanoma skin cancer (n = 29,332), women who underwent a mastectomy prior to baseline (n = 2457), and women with less than one year of follow-up (n = 2150), leaving 239,436 women in the final analysis ( Figure 1 ). Based on an established list of morbidities, which was originally designed by Barnett et al. [ 19 ] to measure multi-morbidity in a large population-based dataset and subsequently validated in the UK Biobank cohort ( Appendix A , Table A1 ) [ 29 ], we defined 35 morbidities based on baseline self-reported health conditions ( Figure 2 ). For each woman, we computed the total number of morbidities and categorized them as none/one/multi-morbidity (at least two morbidities). We defined breast cancer as a diagnosis of invasive or in situ breast cancer, using the international classification of diseases (ICD) versions 9 and 10 (ICD-10: C50 or D05; ICD-9: 174 or 2330). We considered only breast cancer cases that were the first cancer diagnosed. All confounding factors (age at menarche, age at menopause, menopausal hormone therapy use, oral contraceptive use, parity and age at first birth, body mass index (BMI), ethnicity, Townsend score, level of physical activity, alcohol consumption) were measured/collected at baseline. We selected well-established breast cancer risk factors based on previous studies [ 30 , 31 ]. We also selected variables that were statistically significantly associated with both morbidity and breast cancer risk ( p -value < 0.05) as confounders if their inclusion in the age-adjusted Cox models changed the hazard ratio by 5% or more [ 32 ]. See Appendix A , Table A2 for more details on the variables of interest, their definition, and information sources. Among 35 baseline self-report morbidities, we included only morbidities with a prevalence of more than 1% ( Figure 2 ) to obtain stable clustering results [ 33 ]. We used MCA [ 34 , 35 ] and cluster analysis to identify morbidity patterns. MCA can produce the input data for the cluster analysis, while reducing noise by excluding unnecessary dimensions that do not contribute significantly to the cluster’s classification. We determined the optimal number of dimensions to extract based on the elbow rule in the Scree plot [ 34 ] and Horn’s parallel analysis for common factor analysis [ 36 ]. Using the numerical outputs of the MCA, we performed agglomerative hierarchical clustering (AHC) preceded by K-means clustering with 2000 initial cluster seeds [ 37 ], through the HCPC function of the Factominer package in R. This method allowed us to reduce the required memory allocations [ 38 , 39 ]. We considered the distance between points in Euclidean space as the distance metric [ 40 ], and Ward’s method was used to create homogeneous clusters by fusion [ 36 ]. We chose the optimal number of clusters, i.e., the identified morbidity patterns and assessed cluster quality, using the Davies–Bouldin [ 41 ] and the GAP indexes [ 42 ]. The optimal number of clusters was the one that corresponded to the minimum value of the Davies–Bouldin index and to the maximum Gap statistics index. Within each cluster, we computed the observed/expected ratios (“O/E-ratios”) for each single morbidity, i.e., the ratio between the prevalence of a given condition in a cluster and its prevalence in the overall study population. Similarly, we computed the exclusivity of each single morbidity, i.e., the number of individuals that had a given morbidity in a cluster over the number of individuals with the same morbidity in the whole study population. A morbidity was considered part of a given morbidity cluster when its O/E-ratio was ≥2 and its exclusivity was ≥25% [ 23 , 43 ]. We named the morbidity patterns based on the predominant morbidities in the clusters. The follow-up time started at the date of first registration at a UK Biobank center and ended at the date of the first cancer diagnosis (any cancer diagnosis, except non-melanoma skin cancer) or mastectomy, death, loss to follow-up, or 31 March 2016, whichever came first. We fitted Cox proportional hazard models to estimate hazard ratios and 95% confidence intervals (95%CIs) of breast cancer risk associated with each single pre-existing baseline morbidity included in the cluster analysis, the number of morbidities, and the morbidity patterns. The timescale was the follow-up time. We graphically assessed the proportional hazards assumption using scaled Schoenfeld residuals plots and log linearity assumption (for quantitative covariates) using Martingale residuals plots and deviance residuals plots. The final multivariable Cox models were adjusted for age at baseline, age at menarche, age at menopause, menopausal hormone therapy use, oral contraception use, parity and age at first birth, BMI, ethnicity, the Townsend score, level of physical activity, and alcohol consumption. We tested the modifying effects of age at baseline, the adherence to the recommendations for breast cancer screening, the BMI, the socioeconomic status, the physical activity level, and the menopause status at baseline with the likelihood ratio test. We conducted several sensitivity analyses: (i) we restricted analyses to menopausal women; (ii) we considered only invasive breast cancer as the outcome; (iii) we used the attained age as the timescale; (iv) we considered death and diagnosis of non-breast cancer as competing risks, using sub-distribution hazards models [ 44 ]; (v) we extracted 11 MCA dimensions, which accounted for more than 70% of the total variability among the study population, as recommended by Higgs [ 45 ]; we also extracted all dimensions, assuming they were all significant, and kept different numbers of clusters (3 and 4 clusters) with both 11 and all dimensions extracted. All statistical analyses were performed using R version 4.1.0.

Breast cancer is the most common female cancer, with 2,088,849 new cases worldwide in 2018, accounting for 11.6% of incident cancer cases [ 1 ]. Despite decades of intensive research effort, only about 70% of the disease occurrence is explained by well-established risk factors [ 2 ]. Most of the identified risk factors are not readily modifiable [ 2 , 3 , 4 ], leading to a need for additional research to better understand etiologic processes. In developed countries, most breast cancer cases are diagnosed among women of middle age or older [ 5 ], coinciding with the occurrence of other long-term morbidities [ 6 , 7 ]. Previous studies have suggested associations between breast cancer risk and specific chronic diseases, such as endocrine disorders [ 8 , 9 ], inflammatory conditions [ 10 ], autoimmune diseases [ 11 ], and cardiovascular diseases (CVDs) [ 12 ], especially among postmenopausal women. The underlying mechanisms of these associations could involve shared common physiopathological pathways (e.g., estrogen-related pathways, inflammation pathways) [ 13 , 14 , 15 , 16 ], shared genetic predispositions, shared risk factors (e.g., obesity, physical inactivity) [ 17 ], and medications (e.g., aspirin) [ 18 ]. As people get older, they often develop two or more chronic diseases. With an aging population, the number of people experiencing several multi-morbidities is rising globally [ 6 , 7 , 19 , 20 , 21 ]. In the general population, co-existing morbidities could be classified into common clinically meaningful patterns [ 22 , 23 ]. Sharing underlying biological mechanisms and/or sets of risk factors, the morbidities in the same cluster often interact mutually, which complicates treatments and management and increases the risk of adverse events above and beyond the sum of the risk of individual disease [ 24 ]. Being diagnosed with multi-morbidity is also associated with an increased likelihood of being subjected to breast cancer screening [ 25 , 26 , 27 ], which may lead to increased surveillance of breast cancer incidence. Thus, it is necessary to consider patterns of morbidity, in addition to associations with single chronic diseases, with breast cancer risk. However, to date, there is no epidemiological evidence as to whether and to what extent breast cancer risk varies according to different patterns of morbidity. In this context, our study aimed to investigate the association between the number of morbidities and patterns of morbidity and the risk of female breast cancer.

In the study population, the median age at baseline was 57.7 years (interquartile range [IQR]: 50.2, 63.2). At least one morbidity was present in 58.4% of women at baseline, and the prevalence of multi-morbidity was 25.8%. Hypertension was the most prevalent morbidity (23.1%), followed by painful conditions (17.2%) and asthma (12.3%). The prevalence of obesity was 23.5%, and 23.5% of women had menopause after the age of 51 at baseline. Most women were postmenopausal (73.5%) and were adherent to breast cancer screening recommendations (66.6%) at baseline, as assessed at recruitment ( Table 1 , Figure 2 ). During a median follow-up time of 7.1 years (IQR: 6.4, 7.8), 5,326 women developed breast cancer (2.0%). We considered the first five MCA dimensions (see Supplementary, Figures S1–S3 ), which explained 39% of the total variance, as input to the clustering algorithms. We identified five baseline morbidity patterns (see Supplementary, Figures S4 and S5 & Table S1 ), named as follows: Pattern 1—no-predominant morbidity, pattern 2—psychiatric morbidities, pattern 3—respiratory/immunological morbidities, pattern 4—cardiovascular/metabolic morbidities, pattern 5—unspecific morbidities (see Table 1 ). The median age at baseline was 57.4 years (IQR: 49.9, 63.0), and the median follow-up time was 7.1 years (IQR: 6.4, 7.8). There was no morbidity with an O/E ratio ≥ 2. The main features of this pattern were the low rate of multi-morbidity (6.9%) and the high rate of the absence of morbidity (62.6%). The median age at baseline was 55.7 years (IQR: 48.7, 61.7), and the median follow-up time was 7.0 years (IQR: 6.3, 7.8). Women with this pattern were predominantly diagnosed with anxiety and depression disorders. The median age at baseline was 56.7 years (IQR: 49.1, 62.8), and the median follow-up time was 7.1 years (IQR: 6.4, 7.8). Women with this pattern were predominantly diagnosed with psoriasis/eczema, COPD, and asthma. The median age at baseline was 62.6 years (IQR: 57.2, 66.4), and the median follow-up time was 7.0 years (IQR: 6.3, 7.8). Women with this pattern were predominantly diagnosed with diabetes, stroke, and coronary–heart disease. The main features of this pattern were the high proportions of elderly (about 65% were 65 years or older at baseline), multi-morbidity (96.6%), and deprived people (37.1% of women with this pattern were in the quintile with the highest levels of deprivation). The median age at baseline was 59.2 years (IQR: 51.9, 64.0), and the median follow-up time was 7.1 years (IQR: 6.4, 7.8). Women with this pattern were predominantly diagnosed with migraine, diverticular intestine disease, inflammatory bowel disease, rheumatoid disease, and threated dyspepsia. In both age-adjusted and fully adjusted models, no significant association was found between either the number of morbidities or any morbidity pattern and breast cancer risk, but there was a 12% increased risk associated with self-reported depression ( Table 2 and Table 3 ). The results did not vary significantly with age at baseline ( p -value interaction = 0.43 and 0.07, for the analyses on the number of morbidities and morbidity patterns, respectively) and adherence to recommendations for breast cancer screening among women aged 50 and older ( p -value interaction = 0.44 and 0.84, for the analyses on the number of morbidities and morbidity patterns, respectively), although we found an increased risk among women aged of up to 50 years in the psychiatric morbidities pattern (HR= 1.25; 95%CI: 1.02–1.52) ( Figure 3 and Figure 4 ). The results remained consistent after accounting for competing risks ( Table 4 ), when considering attained age as the timescale in the Cox models ( Supplementary, Table S2 ) and in other sensitivity analyses (see Supplementary, Tables S3 and S4 ).

Among female participants in the UK Biobank cohort, 58.4% had at least one chronic disease, while 25.8% had two or more simultaneous morbidities. Hypertension was the most prevalent disease (23.1%) at baseline. We found five morbidity patterns: no-predominant morbidity, psychiatric morbidities, respiratory/immunological morbidities, cardiovascular/metabolic morbidities, and unspecific morbidities. There was a 1.12-fold increased risk among women who self-reported depression and a 25% increased risk of breast cancer associated with a psychiatric morbidity pattern compared to that with the no-predominant morbidity pattern, among women younger than 50 only. We did not observe other significant associations between either the number of morbidities or any morbidity pattern and the risk of breast cancer, which did not vary according to adherence to breast cancer screening recommendations, socioeconomic status, BMI, physical activity level, or menopausal status. Despite heterogeneous findings in previous studies on morbidities across different populations and settings, several morbidity patterns often emerge in the literature, which were also observed in our study [ 22 , 23 , 46 , 47 ]. The pattern of cardiovascular/metabolic morbidities has been extensively described previously, as there are established etiologic associations among diabetes, stroke, heart failure, and heart disease, with an interlinked pathophysiology and common risk factors, such as obesity, physical inactivity, and smoking [ 48 ]. For the pattern of psychiatric morbidities, although little is known about the pathogenesis of depression and anxiety, these two frequent mental illnesses share a largely overlapping set of risk factors with breast cancer, including female sex, genetic predisposition, family history, and environmental influence (childhood adversity, low socioeconomic status) [ 49 , 50 ]. Depression and anxiety are also common coexisting conditions among patients with chronic comorbidities, including cancer [ 51 , 52 ]. Consistent with our findings, a recent nationwide population-based study has shown that mental disorders were associated with a subsequent higher risk of cancer, although the causal link remains a topic of debate [ 52 ]. The diseases included in the respiratory pattern, such as chronic obstructive pulmonary disease and asthma, involve a prolonged inflammatory response and the sharing of risk factors, such as smoking, an unhealthy diet, physical inactivity, and high alcohol consumption. However, combinations among asthma, COPD, and psoriasis and eczema are less common. Thus, these patterns found in our clustering analysis not only represent a clinically relevant morbidity status in women in the UK Biobank cohort but also reflect distinct profiles of (known or unknown), shared genetics, and behavioral and environmental risk factors, both of which might increase the risk of developing cancer. Indeed, to our knowledge, our study is the first to investigate the association between morbidity patterns and breast cancer risk. We found no association between either the number of morbidities or morbidity patterns and breast cancer risk, regardless of the women’s age at baseline, and socioeconomic characteristics, apart from an increased risk among women aged less than 50 having multiple psychiatric diseases. Analyses stratified based on adherence to breast cancer screening recommendations did not modify our main results, suggesting that surveillance bias is not an important modifying factor in the association between breast cancer risk and morbidities. Previously, there was only a case-control study reporting results on the association between multi-morbidity and breast cancer risk. The findings indicated that an increasing number of morbidities measured with the Charlson comorbidity index (CCI) was associated with an increasing breast cancer risk (46,324 cases) after a 10-year follow-up of women aged 45–85, but no association was found for individual morbidities [ 53 ]. However, they were not able to control for confounding factors other than age at baseline and to account for surveillance bias. For comparison purposes, we applied the same methods in an additional analysis by using the CCI ( Tables S5 and S6 ), and we did not find a significant association between the Charlson morbidity number and breast cancer risk after adjusting for well-known risk factors. There are several hypotheses to explain the null results. First, women with morbidity could experience other serious long-term outcomes before a breast cancer diagnosis. Indeed, when accounting for death and malignancies other than breast cancer as competing risks, we found that compared to that in women with no predominant morbidity, women with other patterns were more likely to die and/or to be diagnosed with other cancers. This is particularly pronounced among women with cardiovascular/metabolic and respiratory/immunological morbidities. Second, given the different biologic characteristics of divergent breast cancer subtypes [ 54 ] and the complexity of multi-morbidity mechanisms and risk factors, the risk estimations could vary across individual associations, and the possible opposing effects could drive the combined estimates toward null. For instance, BMI, a common risk factor of various morbidities, is strongly associated with hormone receptor-positive tumors, but not a triple-negative or core basal phenotype [ 55 ]. A high BMI is a risk factor of postmenopausal breast cancer, but a protective factor of premenopausal breast cancer. Type 2 diabetes is an independent risk factor of breast cancer risk in postmenopausal women, but no increased risk was observed for premenopausal women [ 8 ]. In our study, when restricting analyses to postmenopausal women only, the null associations remained consistent. Previous large prospective cohorts reported that low socioeconomic positions, a contributing factor of psychiatric morbidities, were found to be associated with a lower risk of ER+ breast cancer but a higher risk of the ER- subtype [ 56 , 57 ]. Meanwhile, adverse life events, such as childhood abuse and divorce, were associated with a higher risk of ER+, but not ER-, breast cancer [ 57 , 58 ]. Third, our null results could also suggest that the underlying common biological pathways among morbidities in an individual pattern and their shared risk factors were not a key factor explaining breast cancer risk after accounting for established breast cancer risk factors. Strength and limitations: The UK Biobank cohort is a large population-based cohort with a high follow-up rate and important number of breast cancer cases. The cohort includes a wide range of information on personal medical history, reproductive factors, lifestyle factors, socioeconomic status, and family medical history, with low levels of missing data. Nevertheless, there are several limitations that must be acknowledged. Assuming that the prevalence of having at least one morbidity in women in the UK Biobank cohort is slightly lower than what has been found (42.2%, 33.8%) in previous studies of Barnett and Gondek, respectively (since these studies have included data of both women and men in the analyses, which could lead to a potential underestimation of the morbidity prevalence), this suggests the occurrence of “healthy” volunteer bias (i.e., UK Biobank participants are more likely to be in good health conditions than the general population) [ 59 , 60 ]. However, since our study focuses on investigating breast cancer risk in relation to morbidity and not on estimating disease prevalence rates and many people with a wide range of morbidities and risk factors are included in the cohort, the risk estimations are unlikely to be biased [ 59 , 60 ]. We used self-reported health condition data, which were not externally validated, and the UK Biobank did not include information on morbidity severity. There was no longitudinal updated morbidity status and thus no possibility to study changes in morbidity patterns during follow-up. We also missed details on the breast cancer stage, grade, and receptor status. This did not allow us to further study the surveillance biases related to the disease stage and grade or to investigate potential pathways related to tumor receptor status.

Female participants in the UK Biobank cohort can be classified into five morbidity patterns: no-predominant morbidity, psychiatric morbidities, respiratory/immunological morbidities, cardiovascular/metabolic morbidities, and unspecific morbidities. We found a significant increased risk among women aged younger than 50 with a psychiatric diseases pattern, but there was no other significant association among the number of morbidities, the morbidity patterns, and the risk of breast cancer in this population. Our findings suggest that multimorbidity is not a decisive factor to help identify patients at increased risk of breast cancer.