Age-Adjusted Phenome-Wide Analysis Identifies Liver Enzymes and Lipid Markers Associated with Prevalent Breast Cancer in the Qatar Biobank Authors: Lama Abujamous, Shereena Mohd Arif, Hamda Al-Thawadi, Rozaimi Razali Published: 2026 ABSTRACT Most evidence linking phenotypic traits to breast cancer risk derives from Western populations, leaving a critical gap for the Middle East and North Africa region, where breast cancer presents at younger ages and where demographic, metabolic, and lifestyle profiles differ substantially. To address this, we conducted an exploratory phenome-wide association analysis within the Qatar Biobank, systematically screening 93 demographic, anthropometric, biochemical, lifestyle, and medical history variables among 96 women with physician-confirmed breast cancer and 471 healthy controls. A three-stage filtering pipeline: i) Benjamini–Hochberg-corrected significance, ii) Cliff''s Delta effect sizes, and iii) crude odds ratios, reduced 42 initial associations to 18 candidates, which were then evaluated using age-adjusted logistic regression with confounding quantified by percentage change in log-transformed odds ratios. Age was the dominant correlate of breast cancer status (OR = 1.10 per year, 95% CI: 1.08–1.13), and its adjustment fundamentally reshaped the association landscape: metabolic and anthropometric markers including glycated haemoglobin, body mass index, and waist circumference lost significance entirely, with effect attenuations exceeding 100% and reversal of effect direction, revealing these associations as artefacts of the 20-year median age gap between groups. In contrast, alanine aminotransferase, aspartate aminotransferase, total cholesterol, and LDL cholesterol remained statistically significant after adjustment for age. Multivariable models indicated that these four markers correspond to two mutually independent associations, hepatic and lipid, neither explained by adiposity or glycaemic status. Because all measurements were taken at a single visit after diagnosis had already occurred in the case group, these differences cannot be assigned a temporal direction and may reflect the consequences of disease, its treatment, or selective survival rather than pre-diagnostic biology. They are reported as exploratory cross-sectional correlates of prevalent breast cancer status requiring replication in an incident-case cohort, not as risk factors.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s44197-026-00628-6. Background Cancer remains one of the leading causes of death worldwide, with breast cancer being the most commonly diagnosed malignancy and a major cause of mortality among women [1]. In 2020, approximately 2.3 million new cases were recorded globally [2, 3]. Breast cancer risk is shaped by a complex interplay of non-modifiable factors (advancing age, family history, and inherited susceptibility) and modifiable exposures including obesity, physical inactivity, and dietary patterns [4–6]. Increased body fat, particularly central adiposity, has been strongly associated with postmenopausal breast cancer through effects on estrogen metabolism and chronic inflammation [7, 8]. Dysregulation in lipid metabolism, particularly elevated low-density lipoprotein cholesterol (LDL-C), has also been implicated through pro-inflammatory and hormonal pathways [9], while vitamin D deficiency has been linked to poorer breast cancer outcomes via roles in cell differentiation and immune regulation [10, 11]. Age-related metabolic changes further modulate these associations, with postmenopausal obesity serving as a major contributor through its effects on estrogen and insulin pathways [12, 13]. However, the majority of evidence on these associations derives from Western populations. In the Middle East and North Africa (MENA) region, breast cancer often presents at a younger median age (45–51 years) compared with Western cohorts (typically > 60 years), a pattern attributed to differences in reproductive factors, rapid lifestyle westernisation, and possibly genetic predisposition [14–16]. Population-specific phenotypic associations remain understudied in this region, limiting the generalisability of established risk prediction models and screening strategies to Middle Eastern populations. Biobanks such as the Qatar Biobank (QBB) provide a population-based infrastructure for systematically linking questionnaire data, physical measurements, and laboratory markers with health outcomes. QBB enrols well-characterised participants using standardised clinic protocols and allows linkage to electronic medical records (EMR), enabling efficient identification of individuals with major chronic diseases, including breast cancer [17]. For Middle Eastern populations, which remain under-represented in large international cohorts, QBB offers a unique opportunity to generate region-specific evidence on phenotypic associations and to examine whether patterns observed in Western settings apply in this context. A phenome-wide association study (PheWAS) approach is well suited to this objective, as it enables hypothesis-free identification of associations that may be missed by candidate-variable studies. In this study, we conducted an exploratory phenome-wide investigation comparing confirmed breast cancer cases with healthy controls drawn from QBB, screening 93 demographic, anthropometric, biochemical, lifestyle, and medical history variables. This represents, to our knowledge, the first comprehensive phenotypic screen in relation to breast cancer status in a Middle Eastern biobank cohort. Methods Ethics Approval This study was conducted according to the ethical guidelines outlined by (QBB) and Qatar’s national research regulatory frameworks. Ethical approval for the use of QBB phenotype data was obtained from the IRB number QF-QBB-RES-ACC-00211. Data Source The phenotype data employed in this study were obtained from QBB, a national population-based cohort in Qatar. QBB collects an extensive range of observable traits from enrolled participants, encompassing both biological and behavioural dimensions. Variables span demographics (age, gender, nationality), anthropometry (height, weight, BMI, waist-to-hip ratio), clinical and biochemical indices (blood pressure, lipids, renal and liver markers, vitamin D), lifestyle factors (smoking, activity, diet), self-reported medical and family histories, disease status, medication use and outcomes. All measures were collected with uniform protocols during clinic visits, and disease status can be ascertained through linkage to electronic medical records (EMR) [17]. The data used in this analysis reflect values recorded at the participant''s enrolment visit. The QBB extract provided for this study is cross-sectional: all phenotypic measurements derive from a single enrolment visit, with no longitudinal follow-up. Breast cancer status was ascertained retrospectively through EMR linkage, so all cases are prevalent rather than incident. Study Population From approximately 14,000 Qatar Biobank participants with available phenotype data at the time of extraction, we restricted the analytic sample to female participants aged ≥ 18 years. Participants were then classified into two groups based on the following criteria. The confirmed breast cancer cases (CC) were identified through linkage to participants'' electronic medical records (EMR), requiring a physician-documented breast cancer diagnosis recorded prior to or at the time of QBB enrolment. Linkage was to Hamad Medical Corporation, Qatar''s national tertiary referral hospital. The EMR extract records the presence of a physician-documented breast cancer diagnosis but does not distinguish a primary breast malignancy from secondary involvement of the breast by another primary tumour, and does not record laterality, stage, or grade. Because diagnoses predate or coincide with enrolment, the design does not permit temporal ordering of exposure and outcome. Participants were excluded from the CC group if they had a documented history of other malignancies, missing or inconsistent EMR information regarding breast cancer diagnosis, or incomplete data for key phenotypic variables. This process yielded 96 confirmed cases. The healthy controls (HC) were drawn from the remaining female participants who met all the following criteria: (i) no physician-documented diagnosis of any disease in the EMR, and (ii) no self-reported disease history via the QBB health questionnaire. No sampling, frequency matching, or restriction beyond these criteria was applied; all 471 participants meeting the HC definition were included. The resulting case-to-control ratio of approximately 1:4.9 reflects the natural composition of the eligible QBB cohort rather than a pre-specified design target. Because QBB is a population-based biobank and not a cancer registry, the number of confirmed cases reflects the relatively low prevalence of breast cancer in the underlying cohort at the time of data freeze. This modest case sample size reinforces the exploratory, hypothesis-generating nature of the present analysis. Variable Selection A total of 93 phenotypic variables were included in the analysis, spanning demographics (age, nationality), anthropometry (height, weight, BMI, waist and hip circumference, waist-to-hip ratio), clinical biochemistry (lipid panel, liver enzymes, renal markers, glycaemic indices, hormones, vitamins, haematological parameters), body composition (bone mineral density and content), lifestyle factors (smoking, physical activity, dietary variables), and self-reported medical and family history items from the QBB health questionnaire. Variables were selected based on availability in the QBB phenotype dataset and prior literature suggesting plausible associations with breast cancer risk or metabolic health. Four self-reported questionnaire items from the QBB MainQuestionnaire are referred to throughout by their field codes: MQ_H36 (age at diabetes diagnosis, years), MQ_H43 (age at hypertension diagnosis, years), MQ_H40 (age at hypercholesterolaemia diagnosis, years), and MQ_F_Sib_B (number of biological siblings). Responses of "Do not remember," "Do not know," and "Prefer not to answer" were treated as missing (NA). Missing Data Responses of "Do not remember," "Do not know," and "Prefer not to answer" in questionnaire fields, and empty strings in categorical fields, were recoded as missing (NA); character-coded variables were coerced to numeric, with non-convertible entries treated as NA. No variable was excluded on the basis of its missingness rate, and no imputation of any kind was performed. All analyses are complete-case within each test: every univariate comparison and every regression model uses all participants with non-missing values on the variables entering that specific test, so the analytic denominator varies between variables. For the multivariable comparisons, the corresponding age-adjusted reference model was refitted on the identical complete-case subset, so that differences between nested models reflect the added covariate rather than a change in sample composition. Complete-case analysis is unbiased when data are missing completely at random, and more generally when the probability of being a complete case does not depend on the outcome conditional on the covariates in the model [18]. We did not assume this holds uniformly across a 93-variable screen, and two considerations governed the decision not to impute. First, the four markers underpinning the primary findings (ALT, AST, total cholesterol, LDL cholesterol) have ≤ 0.4% missingness overall and none among cases, so no imputation scheme could materially alter these estimates. Second, the assays with substantial missingness show missingness that is differential by group (estradiol 31.2% in CC vs 7.0% in HC,SHBG 53.1% vs 31.0%), which makes a missing-at-random assumption difficult to defend for precisely those variables; imputation under MAR would then impose model-based structure on a hypothesis-free screen and could introduce bias rather than remove it [18, 19]. Those markers are reported descriptively and are not advanced as findings. Per-group missingness rates for all 93 variables are given in Supplementary Table 2. Statistical Analysis All analyses were conducted in R (version 4.4.1) using RStudio (version 2024.12.0 + 467). Two different multiplicity corrections are used deliberately, for two different purposes. For the phenome-wide screen the Benjamini–Hochberg procedure controls the false discovery rate, which is the appropriate criterion for a discovery analysis whose output is a ranked set of candidate hypotheses: the cost of a false negative is high relative to the cost of a false positive, which subsequent studies can correct. Family-wise error control across 93 variables would be inappropriately conservative in this setting. For the minimum detectable effect size, by contrast, the objective is the opposite: to characterise the study''s sensitivity under the least favourable multiplicity assumption. The Bonferroni-corrected α therefore provides a deliberately conservative upper bound on the detectable effect, rather than a decision threshold applied to any reported result. Both the nominal α = 0.05 and the Bonferroni MDES values are reported (d ≈ 0.31 and d ≈ 0.48 respectively) so that the plausible range of study sensitivity is visible. No hypothesis test in this manuscript is evaluated against a Bonferroni threshold. Minimum Detectable Effect Size Given the modest case count, we report the minimum detectable effect size (MDES) rather than observed (post-hoc) power, the latter being statistically uninformative as it is a deterministic function of the observed p-value [20]. For the realised sample (96 cases, 471 controls), the smallest standardised mean difference (Cohen''s d) detectable with 80% power was obtained analytically (two-sample framework,pwr package) at the nominal α = 0.05 and at a Bonferroni-corrected α = 0.05/93 reflecting the 93 screened variables, and expressed as an odds ratio per standard deviation via ln(OR) = 1.81d [21]. A Monte Carlo simulation (1000 replicates) corroborated these values under age-adjusted logistic regression using the empirical age distribution. Descriptive Statistics and Distributional Assessment Central tendency and variability were summarised for each variable (mean, SD, median, IQR, range, and mode; Supplementary Table 1). Distributional normality was assessed using the Kolmogorov–Smirnov test. Because the majority of variables deviated significantly from normality (p < 0.05), non-parametric methods were used for all univariate comparisons. Univariate Screening Group differences between CC and HC were evaluated using the Mann–Whitney U test. Raw p-values were adjusted for multiple comparisons using the Benjamini–Hochberg (BH) procedure to control the false discovery rate (FDR). Effect sizes were quantified using Cliff''s Delta, with magnitudes interpreted as negligible (|δ|< 0.147), small (0.147 ≤|δ|< 0.33), medium (0.33 ≤|δ|< 0.474), and large (|δ|≥ 0.474) following the thresholds proposed by Romano et al. [22]. Inter-variable correlations were assessed using Spearman''s rank correlation. Univariate Odds Ratios To complement the non-parametric tests with an estimate of association magnitude, crude odds ratios were computed for each variable. Continuous variables were dichotomised at the pooled-sample median (CC and HC combined) into "high" (above median) and "low" (at or below median) categories, and ORs with 95% confidence intervals were obtained from 2 × 2 contingency tables using Fisher''s exact test. This median-split approach was adopted as an exploratory screening heuristic to provide a common, data-driven scale for ranking variables by association strength alongside the BH-adjusted p-values and effect sizes. The OR > 2.0 filtering threshold was selected to prioritise variables showing at least a twofold difference in the odds of high-versus-low classification between groups, a benchmark commonly applied in exploratory biomarker screening to distinguish potentially meaningful associations from modest ones (Haddow et al., 2004). We acknowledge that dichotomising continuous predictors at the median reduces statistical power and, when data-derived cutpoints are used, can introduce bias [23],accordingly, the median split was confined to this univariate screening stage and never entered any inferential model i.e., all age-adjusted analyses use predictors in continuous form. The resulting ORs are sample-specific rather than clinically anchored, served solely as a filtering criterion, and are not the basis for the study''s primary conclusions. Variable Filtering Variables were retained for multivariable modelling if they met all three of the following criteria from the univariate screening: (i) BH-adjusted p-value ≤ 0.05, (ii) Cliff''s Delta |δ|≥ 0.11, a threshold set below the conventional small-effect boundary of 0.147 [22] to retain variables showing any non-trivial group separation in this exploratory screen, and (iii) crude OR > 2.0. The three criteria were applied jointly as a presentation heuristic to rank an initially large candidate set, not as an inferential gate: they determine which variables are displayed in the main-text figures and tables, and no reported conclusion is conditional on them. The OR > 2.0 threshold was chosen because a twofold difference in odds is a conventional benchmark for distinguishing associations of potential practical relevance from marginal ones in exploratory biomarker screens (Haddow et al., 2004),we acknowledge it is a convention rather than a principled cut-point, as are the Cliff''s δ and BH thresholds. To demonstrate that the conclusions do not depend on this choice, age-adjusted models were additionally fitted for every variable surviving FDR correction with both the Cliff''s δ and odds-ratio filters removed. Of the 34 variables passing FDR correction among the 70 continuous candidates, only ALT, AST, total cholesterol, and LDL cholesterol retained significance after age adjustment and within-panel BH correction, which is identical to the filtered analysis (Supplementary Table 4). The filter therefore changed which variables were foregrounded, not which were concluded. Application of these filters reduced the initial 42 nominally significant variables (raw p ≤ 0.05) to 18 candidates carried forward to age-adjusted modelling. Because a per-unit logistic predictor requires a continuous scale, the candidate set for this analysis comprised the 70 numeric-coercible variables with sufficient spread (≥ 30 non-missing values, ≥ 10 distinct levels), excluding identifiers, the outcome, the age covariate, and categorical questionnaire/dietary items; 34 of these survived Benjamini–Hochberg correction. The three-criterion filter is therefore retained only as an exploratory device for ranking and presenting the main-text variables, not as an inferential selection gate. Full results are reported in Supplementary Table 4. Age-Adjusted Logistic Regression and Confounding Assessment To evaluate whether the associations identified in univariate screening were independent of age we employed a two-model logistic regression strategy for each of the 18 filtered variables. In the first model (crude), each variable was entered as the sole continuous predictor of group status (CC vs. HC) using binary logistic regression. In the second model (age-adjusted), the same variable was modelled jointly with age (in years) as a covariate. In both models, predictor variables were entered in their original continuous form (i.e., without dichotomisation), and odds ratios represent the change in odds per one-unit increase in the predictor. Age was entered as a continuous linear term on the log-odds scale in all adjusted models; it was not categorised, and no spline or polynomial expansion was fitted. This specification was pre-specified on grounds of parsimony and interpretability. With 96 cases distributed across 18 candidate models, a restricted cubic spline expansion would consume two additional degrees of freedom per model and reduce the effective number of events per estimated parameter, a concern given the estimation instability already documented for waist-to-hip ratio. The linear specification also permits the age coefficient to be reported as a single interpretable odds ratio per year of age. Reliance on this assumption is limited by design rather than by test: the age-restricted sensitivity analyses reported in Supplementary Table 3 confine the comparison to a common-support window (19–76 years) and to a narrower balanced window (40–65 years), within which any curvature in the age–outcome relation is substantially attenuated. That ALT, AST, total cholesterol and LDL cholesterol retained significance under both restrictions indicates that these associations do not depend on the functional form assumed for age across the full age range. We nonetheless note that no formal test of departure from linearity was performed, and a non-linear age effect not captured by the linear term would contribute to the residual confounding acknowledged in the Limitations. Confounding by age was quantified as the absolute percentage change between the crude and age-adjusted log-transformed odds ratios: |log(OR_crude) − log(OR_adjusted)|/|log(OR_crude)|× 100. Variables showing ≥ 10% change were considered meaningfully confounded by age, following established epidemiological conventions [24, 25]. Statistical significance in the age-adjusted models was assessed after applying BH correction to control the FDR at 0.05. Menopausal status was not recorded in the QBB research extract, and the assay panel did not include gonadotrophins (follicle-stimulating and luteinising hormone) that would permit biochemical staging. Circulating estradiol was not used as a surrogate, as it is confounded by exogenous and antineoplastic endocrine therapy and was differentially missing between groups (31.2% in cases vs 7.0% in controls). Age, the principal determinant of menopausal status, was included as a covariate in all adjusted models. Multivariable models were additionally fitted for the four markers retaining age-independent significance. Each was adjusted for age together with BMI, waist circumference, or HbA1c to assess confounding by adiposity and glycaemic status; markers were then entered pairwise and jointly to assess mutual independence. Collinearity was quantified by the variance inflation factor, computed as 1/(1 − R2ⱼ) from the regression of each predictor on the remaining predictors. Nested models were compared by likelihood-ratio test. Benjamini–Hochberg correction was applied within each panel to marker terms only, adjustment covariates being excluded from the correction family. Complete-case subsets were matched across nested models as described under Missing data. Model Diagnostics and Separation Assessment For each age-adjusted logistic regression model, we examined whether the maximum likelihood estimation had converged and inspected for signs of complete or quasi-complete separation. This is a condition in which a predictor perfectly or nearly perfectly discriminates between outcome groups, producing inflated coefficient estimates and extremely wide confidence intervals [26, 27]. Separation was suspected when the estimated odds ratio exceeded 10 and the 95% confidence interval spanned more than two orders of magnitude. Where such instability was identified, results are reported but flagged explicitly, and the corresponding point estimates should be interpreted with caution. For the single model flagged with quasi-complete separation (waist-to-hip ratio), we refitted the age-adjusted model using Firth''s penalised-likelihood logistic regression (logistf), which guarantees finite, first-order bias-reduced estimates with profile penalised-likelihood confidence intervals [28]. Because waist-to-hip ratio is bounded within a narrow range, the Firth estimate is additionally expressed per 0.1-unit increment for interpretability. Maximum-likelihood estimates are retained for all non-separated predictors. Age-restricted Sensitivity Analysis To address the case–control age imbalance beyond covariate adjustment, age-adjusted models were refit on the case–control common-support age range and on a narrower, age-balanced 40–65-year window (the two windows and the group age distributions are shown in Supplementary Fig. 1). Age matching and propensity-score approaches were not used: with 96 cases and limited age overlap they produce unstable estimates and, in the case of matching, discard observations and reduce power without resolving the cross-sectional, prevalent-case design. Clinical Reference-range Sensitivity Analysis To assess whether the associations retained after age adjustment are clinically actionable rather than artefacts of continuous parameterisation, the four markers retaining age-independent significance (ALT, AST, total cholesterol, LDL cholesterol) were dichotomised at established clinical thresholds and re-examined by age-adjusted logistic regression (abnormal vs normal; age as a continuous covariate). ALT was dichotomised at the traditional upper reference limit (≥ 33 U/L) as the primary threshold and, given the all-female cohort, additionally at the women-specific limit of 19 U/L [29] as a secondary sensitivity analysis. Remaining thresholds were AST ≥ 32 U/L (laboratory upper reference limit), total cholesterol ≥ 5.2 mmol/L and LDL cholesterol ≥ 3.4 mmol/L [30]. p-values were Benjamini–Hochberg-adjusted across the four primary tests,the secondary ALT threshold is reported with its unadjusted p. Results Cohort Characteristics Participants were stratified into HC (n = 471) and CC (n = 96). The participant selection process is illustrated in Fig. 1. The CC group was substantially older (median 53 vs. 33 years) and showed higher BMI, waist circumference, and waist-to-hip ratio compared with HC. Data completeness was high for the variables underpinning the primary findings: ALT, AST, and total cholesterol were each available for 566 of 567 participants (0.2% missing) and LDL cholesterol for 565 (0.4%), with no missing values among cases for any of the four. Across the biochemical panel most variables had under 5% missingness; the principal exceptions were myoglobin (43.7%), C-reactive protein (41.6%), and the sex-hormone assays sex-hormone-binding globulin (34.7%) and estradiol (11.1%). Notably, missingness in the hormonal assays was differential by group (estradiol 31.2% in CC vs 7.0% in HC; SHBG 53.1% vs 31.0%), which together with their treatment-confounded interpretation, provides a further reason these markers are not advanced as findings. By contrast, the four markers retained after age adjustment had negligible, non-differential missingness, so the corresponding complete-case estimates are not materially affected. Per-group missingness for all 93 variables is reported in Supplementary Table 2.Fig. 1Participant selection and analytic workflow. From approximately 14,000 Qatar Biobank participants with phenotype data, female participants aged ≥ 18 years were retained (male participants excluded) and classified by electronic medical record (EMR) linkage and self-reported history into CC (n = 96; physician-documented diagnosis via EMR linkage) and HC (n = 471; no EMR-documented diagnosis and no self-reported disease history). Participants with other malignancies, missing or inconsistent EMR data, or incomplete key variables were excluded. The resulting analytic sample (N = 567) was screened across 93 phenotypic variables using Mann–Whitney U tests with Benjamini–Hochberg (BH) correction, Cliff''s delta, and crude odds ratios. Eighteen variables meeting all three filter criteria (BH-adjusted p ≤ 0.05, |Cliff''s δ|≥ 0.11, crude OR > 2.0) advanced to crude versus age-adjusted logistic regression; five (age, ALT, AST, total cholesterol, LDL cholesterol) remained independently associated with breast cancer status after age adjustment Participant selection and analytic workflow. From approximately 14,000 Qatar Biobank participants with phenotype data, female participants aged ≥ 18 years were retained (male participants excluded) and classified by electronic medical record (EMR) linkage and self-reported history into CC (n = 96; physician-documented diagnosis via EMR linkage) and HC (n = 471; no EMR-documented diagnosis and no self-reported disease history). Participants with other malignancies, missing or inconsistent EMR data, or incomplete key variables were excluded. The resulting analytic sample (N = 567) was screened across 93 phenotypic variables using Mann–Whitney U tests with Benjamini–Hochberg (BH) correction, Cliff''s delta, and crude odds ratios. Eighteen variables meeting all three filter criteria (BH-adjusted p ≤ 0.05, |Cliff''s δ|≥ 0.11, crude OR > 2.0) advanced to crude versus age-adjusted logistic regression; five (age, ALT, AST, total cholesterol, LDL cholesterol) remained independently associated with breast cancer status after age adjustment Minimum Detectable Effect Size At the nominal α = 0.05, this study has 80% power to detect a standardised difference of d ≈ 0.31 (OR ≈ 1.77 per SD); enforcing multiplicity across the 93 screened variables (Bonferroni α = 5.4 × 10⁻4) raised this to d ≈ 0.48 (OR ≈ 2.40 per SD). A Monte Carlo simulation under age-adjusted logistic regression reached 80% power at d ≈ 0.36, consistent with the analytic estimate. This simulation assumed independence between the predictor and age; for biomarkers correlated with age, the variance inflation from age adjustment will raise the detectable effect above this value, so d ≈ 0.36 represents a lower bound. This study was therefore powered only for medium-to-large standardised effects. Univariate Comparisons The Mann–Whitney U test presents 42 variables showing statistically significant differences between two study groups (p ≤ 0.05) (see Table 1). The most highly significant differences were observed in demographic and anthropometric measures, particularly age (p = 9.59 × 10⁻25), waist-to-hip ratio (p = 1.34 × 10⁻11), and waist circumference (p = 7.34 × 10⁻11). Metabolic and glycemic markers also showed strong differentiation, including HbA1c (p = 8.44 × 10⁻11), triglycerides (p = 9.15 × 10⁻⁹), and total cholesterol (p = 3.35 × 10⁻⁶). Hormonal differences were evident in estradiol (p = 1.89 × 10⁻13) and testosterone (p = 4.40 × 10⁻⁶), while vitamin D (p = 3.51 × 10⁻⁷) and homocysteine (p = 5.32 × 10⁻⁷) showed notable differences. Additional significant variables included liver enzymes (ALT, AST, alkaline phosphatase), renal markers (creatinine, urea), lipid parameters (LDL cholesterol), insulin-related measures (C-peptide, insulin, glucose), and body composition indices (BMI, weight, bone mineral density/content). Several self-reported questionnaire items also reached nominal significance, but at p-values close to the 0.05 threshold (raw p = 0.011–0.049) rather than the markedly smaller values observed for demographic and metabolic markers (e.g., age, p = 9.6 × 10⁻25): age at hypertension diagnosis (MQ_H43, p = 0.011), age at diabetes diagnosis (MQ_H36, p = 0.013), age at hypercholesterolaemia diagnosis (MQ_H40, p = 0.043), and number of biological siblings (p = 0.049). Overall, the pattern suggests the study groups differ substantially across metabolic, hormonal, anthropometric, and potentially genetic/familial dimensions. Figure 2 displays the significance profile of all 42 variables reaching nominal significance in the Mann–Whitney U tests. Of these, 18 (teal) satisfied all three filtering criteria and were carried forward to age-adjusted modelling.Table 1Participant characteristics and univariate screening by breast cancer status. Values are median (interquartile range) for HC (n = 471) and CC (n = 96). p-values are unadjusted Mann–Whitney U test p-values. All 42 variables reaching nominal significance (raw p ≤ 0.05) are shown. CC, confirmed breast cancer cases; HC, healthy controls. Variables are ordered alphabeticallyCategoryVariableHC median (IQR)CC median (IQR)Raw pOther biochemistry25(OH) Vitamin D (nmol/L)16.0 (11.0–24.0)23.0 (17.0–30.0)3.51476918E-07DemographicsAge (years)33.0 (26.0–43.0)53.0 (45.8–61.0)9.59062198E-25Other biochemistryAlbumin (g/L)43.0 (41.0–45.0)42.0 (38.0–44.3)0.000637368411Liver enzymesAlkaline phosphatase (U/L)64.0 (54.0–77.0)72.5 (61.0–90.0)3.74682968E-05Liver enzymesALT (U/L)13.0 (10.2–19.0)18.0 (13.8–24.0)3.17135315E-07Liver enzymesAST (U/L)16.0 (14.0–19.0)18.0 (14.0–21.0)0.000920812897Other biochemistryBicarbonate (mmol/L)25.6 (24.0–27.0)26.0 (25.0–28.0)0.014319051688AnthropometryBMI (kg/m2)28.0 (24.0–32.7)31.4 (27.7–35.2)7.77651462E-06AnthropometryBody weight (kg)69.8 (60.4–81.6)78.2 (67.2–88.6)6.42506953E-05Glycaemic markersC-peptide (ng/mL)1.9 (1.4–2.7)2.5 (1.7–3.3)0.000167037974Other biochemistryCreatinine (µmol/L)56.0 (51.0–61.8)59.0 (52.0–66.0)0.00284105438Other biochemistryEstradiol (pmol/L)309.5 (136.2–567.5)80.6 (52.0–161.0)1.89357334E-13Other biochemistryFerritin (ng/mL)20.0 (9.0–38.4)38.0 (13.5–79.0)4.20922046E-06HaematologyFibrinogen (g/L)3.4 (2.9–3.8)3.5 (3.1–4.0)0.023319821472Other biochemistryFree triiodothyronine (pmol/L)4.2 (3.7–4.6)3.9 (3.3–4.6)0.020926604995Glycaemic markersGlucose (mmol/L)4.9 (4.6–5.4)5.3 (4.7–6.6)0.000359781856HaematologyHaemoglobin (g/dL)12.6 (11.7–13.3)12.8 (12.1–13.4)0.045996307394Glycaemic markersHbA1c (%)5.3 (5.1–5.6)5.7 (5.3–6.3)8.44497261E-11AnthropometryHip circumference (cm)106.0 (99.0–115.0)110.0 (104.8–116.5)0.007491733114Other biochemistryHomocysteine (µmol/L)8.0 (6.8–10.0)10.0 (8.0–11.5)5.31620683E-07Glycaemic markersInsulin (µU/mL)9.6 (6.7–15.2)12.1 (8.0–18.7)0.008711834795Lipid panelLDL cholesterol (mmol/L)2.6 (2.2–3.2)3.1 (2.4–3.7)0.000226342235Family history (questionnaire)MQ F_Sib_B, Number of biological siblings3.0 (2.0–5.0)4.0 (3.0–6.0)0.048752137065Medical history (questionnaire)MQ H36, Age at diabetes diagnosis (years)36.5 (26.5–45.0)47.0 (35.0–54.0)0.012702016641Medical history (questionnaire)MQ H40, Age at hypercholesterolaemia diagnosis (years)40.0 (32.0–48.0)49.5 (36.0–50.0)0.043483349469Medical history (questionnaire)MQ H43, Age at hypertension diagnosis (years)42.0 (35.0–48.0)49.0 (45.0–54.0)0.011440995894Other biochemistryMyoglobin (ng/mL)19.0 (15.0–21.0)21.0 (17.0–28.0)0.00254739664Other biochemistryNT-proBNP (pg/mL)32.0 (19.4–53.4)40.1 (21.2–69.6)0.043929832494HaematologyPlatelet count (× 10⁹/L)263.0 (225.0–307.0)246.5 (203.8–295.0)0.026969432856Other biochemistryPotassium (mmol/L)4.3 (4.1–4.5)4.4 (4.2–4.7)0.000280988766HaematologyProthrombin time (s)11.8 (11.3–12.5)11.5 (11.0–12.1)0.002834369466HaematologyRed cell distribution width (%)14.0 (13.4–15.4)14.6 (13.9–15.5)0.032001576336Other biochemistrySodium (mmol/L)140.0 (138.0–141.0)140.0 (139.0–142.0)0.0227629961Body compositionTotal bone mineral content (g)2186.1 (2001.5–2364.5)2070.3 (1917.2–2285.3)0.011322864537Body compositionTotal bone mineral density (g/cm2)1.168 (1.095–1.226)1.131 (1.062–1.214)0.035226976351Lipid panelTotal cholesterol (mmol/L)4.7 (4.2–5.2)5.2 (4.5–5.8)3.34819581E-06Other biochemistryTotal testosterone (nmol/L)1.1 (0.8–1.4)0.8 (0.6–1.1)4.39758272E-06Lipid panelTriglycerides (mmol/L)0.9 (0.7–1.3)1.2 (1.0–1.7)9.14721125E-09Other biochemistryUrea (mmol/L)3.7 (3.1–4.5)4.2 (3.5–5.3)0.000129674791Other biochemistryUric acid (μmol/L)241.0 (206.0–283.8)266.0 (222.5–322.5)0.001777587513AnthropometryWaist circumference (cm)80.0 (72.0–90.0)92.0 (85.8–100.0)7.33789408E-11AnthropometryWaist-to-hip ratio0.75 (0.71–0.81)0.82 (0.77–0.90)1.34011686E-11Fig. 2Univariate screening results for 42 variables with nominally significant Mann–Whitney U tests (raw p ≤ 0.05). The x-axis shows − log₁₀(BH-adjusted p-value); the dashed vertical line marks the BH-adjusted significance threshold (p = 0.05). Teal points indicate the 18 variables that satisfied all three filtering criteria (BH-adjusted p ≤ 0.05, |Cliff''s δ|≥ 0.11, crude OR > 2.0); tan points indicate variables that reached nominal significance but did not pass one or more filters. Variables are ordered by − log₁₀(BH-adjusted p-value) Table 1Participant characteristics and univariate screening by breast cancer status. Values are median (interquartile range) for HC (n = 471) and CC (n = 96). p-values are unadjusted Mann–Whitney U test p-values. All 42 variables reaching nominal significance (raw p ≤ 0.05) are shown. CC, confirmed breast cancer cases; HC, healthy controls. Variables are ordered alphabeticallyCategoryVariableHC median (IQR)CC median (IQR)Raw pOther biochemistry25(OH) Vitamin D (nmol/L)16.0 (11.0–24.0)23.0 (17.0–30.0)3.51476918E-07DemographicsAge (years)33.0 (26.0–43.0)53.0 (45.8–61.0)9.59062198E-25Other biochemistryAlbumin (g/L)43.0 (41.0–45.0)42.0 (38.0–44.3)0.000637368411Liver enzymesAlkaline phosphatase (U/L)64.0 (54.0–77.0)72.5 (61.0–90.0)3.74682968E-05Liver enzymesALT (U/L)13.0 (10.2–19.0)18.0 (13.8–24.0)3.17135315E-07Liver enzymesAST (U/L)16.0 (14.0–19.0)18.0 (14.0–21.0)0.000920812897Other biochemistryBicarbonate (mmol/L)25.6 (24.0–27.0)26.0 (25.0–28.0)0.014319051688AnthropometryBMI (kg/m2)28.0 (24.0–32.7)31.4 (27.7–35.2)7.77651462E-06AnthropometryBody weight (kg)69.8 (60.4–81.6)78.2 (67.2–88.6)6.42506953E-05Glycaemic markersC-peptide (ng/mL)1.9 (1.4–2.7)2.5 (1.7–3.3)0.000167037974Other biochemistryCreatinine (µmol/L)56.0 (51.0–61.8)59.0 (52.0–66.0)0.00284105438Other biochemistryEstradiol (pmol/L)309.5 (136.2–567.5)80.6 (52.0–161.0)1.89357334E-13Other biochemistryFerritin (ng/mL)20.0 (9.0–38.4)38.0 (13.5–79.0)4.20922046E-06HaematologyFibrinogen (g/L)3.4 (2.9–3.8)3.5 (3.1–4.0)0.023319821472Other biochemistryFree triiodothyronine (pmol/L)4.2 (3.7–4.6)3.9 (3.3–4.6)0.020926604995Glycaemic markersGlucose (mmol/L)4.9 (4.6–5.4)5.3 (4.7–6.6)0.000359781856HaematologyHaemoglobin (g/dL)12.6 (11.7–13.3)12.8 (12.1–13.4)0.045996307394Glycaemic markersHbA1c (%)5.3 (5.1–5.6)5.7 (5.3–6.3)8.44497261E-11AnthropometryHip circumference (cm)106.0 (99.0–115.0)110.0 (104.8–116.5)0.007491733114Other biochemistryHomocysteine (µmol/L)8.0 (6.8–10.0)10.0 (8.0–11.5)5.31620683E-07Glycaemic markersInsulin (µU/mL)9.6 (6.7–15.2)12.1 (8.0–18.7)0.008711834795Lipid panelLDL cholesterol (mmol/L)2.6 (2.2–3.2)3.1 (2.4–3.7)0.000226342235Family history (questionnaire)MQ F_Sib_B, Number of biological siblings3.0 (2.0–5.0)4.0 (3.0–6.0)0.048752137065Medical history (questionnaire)MQ H36, Age at diabetes diagnosis (years)36.5 (26.5–45.0)47.0 (35.0–54.0)0.012702016641Medical history (questionnaire)MQ H40, Age at hypercholesterolaemia diagnosis (years)40.0 (32.0–48.0)49.5 (36.0–50.0)0.043483349469Medical history (questionnaire)MQ H43, Age at hypertension diagnosis (years)42.0 (35.0–48.0)49.0 (45.0–54.0)0.011440995894Other biochemistryMyoglobin (ng/mL)19.0 (15.0–21.0)21.0 (17.0–28.0)0.00254739664Other biochemistryNT-proBNP (pg/mL)32.0 (19.4–53.4)40.1 (21.2–69.6)0.043929832494HaematologyPlatelet count (× 10⁹/L)263.0 (225.0–307.0)246.5 (203.8–295.0)0.026969432856Other biochemistryPotassium (mmol/L)4.3 (4.1–4.5)4.4 (4.2–4.7)0.000280988766HaematologyProthrombin time (s)11.8 (11.3–12.5)11.5 (11.0–12.1)0.002834369466HaematologyRed cell distribution width (%)14.0 (13.4–15.4)14.6 (13.9–15.5)0.032001576336Other biochemistrySodium (mmol/L)140.0 (138.0–141.0)140.0 (139.0–142.0)0.0227629961Body compositionTotal bone mineral content (g)2186.1 (2001.5–2364.5)2070.3 (1917.2–2285.3)0.011322864537Body compositionTotal bone mineral density (g/cm2)1.168 (1.095–1.226)1.131 (1.062–1.214)0.035226976351Lipid panelTotal cholesterol (mmol/L)4.7 (4.2–5.2)5.2 (4.5–5.8)3.34819581E-06Other biochemistryTotal testosterone (nmol/L)1.1 (0.8–1.4)0.8 (0.6–1.1)4.39758272E-06Lipid panelTriglycerides (mmol/L)0.9 (0.7–1.3)1.2 (1.0–1.7)9.14721125E-09Other biochemistryUrea (mmol/L)3.7 (3.1–4.5)4.2 (3.5–5.3)0.000129674791Other biochemistryUric acid (μmol/L)241.0 (206.0–283.8)266.0 (222.5–322.5)0.001777587513AnthropometryWaist circumference (cm)80.0 (72.0–90.0)92.0 (85.8–100.0)7.33789408E-11AnthropometryWaist-to-hip ratio0.75 (0.71–0.81)0.82 (0.77–0.90)1.34011686E-11 Participant characteristics and univariate screening by breast cancer status. Values are median (interquartile range) for HC (n = 471) and CC (n = 96). p-values are unadjusted Mann–Whitney U test p-values. All 42 variables reaching nominal significance (raw p ≤ 0.05) are shown. CC, confirmed breast cancer cases; HC, healthy controls. Variables are ordered alphabetically Univariate screening results for 42 variables with nominally significant Mann–Whitney U tests (raw p ≤ 0.05). The x-axis shows − log₁₀(BH-adjusted p-value); the dashed vertical line marks the BH-adjusted significance threshold (p = 0.05). Teal points indicate the 18 variables that satisfied all three filtering criteria (BH-adjusted p ≤ 0.05, |Cliff''s δ|≥ 0.11, crude OR > 2.0); tan points indicate variables that reached nominal significance but did not pass one or more filters. Variables are ordered by − log₁₀(BH-adjusted p-value) Variable Selection The filtering process reduced the original 42 significant variables to 18 that demonstrated statistical significance after multiple testing correction, considering effect sizes and ORs (see Table 2). Age emerged as the most differentiating variable with the largest effect size (ES = 0.66, indicating a large effect) and an odds ratio of 11.9, suggesting that individuals in one group are nearly 12 times more likely to be older than those in the comparison group. Anthropometric measures show consistent associations where waist circumference (ES = 0.42, OR = 5.12), waist-to-hip ratio (ES = 0.44, OR = 4.19), BMI (ES = 0.29, OR = 2.71), and weight (ES = 0.26, OR = 2.29) all demonstrated medium effect sizes. Metabolic and glycemic markers such as HbA1c (ES = 0.42, OR = 3.51) and triglycerides (ES = 0.37, OR = 3.49) showed medium effect sizes with substantial odds ratios, pointing toward differences in glycemic control and lipid metabolism while total cholesterol (ES = 0.30, OR = 2.60), LDL cholesterol (ES = 0.24, OR = 2.11), glucose (ES = 0.23, OR = 2.06), and C-peptide (ES = 0.25, OR = 2.43) further support a metabolic syndrome-like differentiation pattern. Liver function and related markers such as ALT (ES = 0.33, OR = 3.14), AST (ES = 0.21, OR = 2.33), and alkaline phosphatase (ES = 0.27, OR = 2.22) suggest hepatic differences between groups, potentially reflecting non-alcoholic fatty liver disease or metabolic stress commonly associated with obesity and insulin resistance. Vitamin D and homocysteine such as total dihydroxyvitamin D (ES = 0.33, OR = 3.18) and homocysteine (ES = 0.33, OR = 2.23) both showed small-to-medium effect sizes, which may reflect nutritional status, cardiovascular risk factors, or metabolic differences. Finally, medical questionnaire items such as MQ_H43 (ES = 0.45, OR = 3.38) and MQ_H36 (ES = 0.36, OR = 2.63) retained significance, suggesting that specific self-reported health conditions or family history factors contribute to group differentiation.Table 2Eighteen variables retained after three-criterion filtering of univariate screening results. ES: Cliff''s Delta effect size (positive values indicate higher values in CC). OR: crude odds ratio from Fisher''s exact test after median-split dichotomisation (see Methods). Variables were retained if BH-adjusted p ≤ 0.05, |Cliff''s δ|≥ 0.11, and crude OR > 2.0. Variables are ordered by descending crude OR. Abbreviations as in Table 1FieldBH-adj pCliff''s δ|ORAge (years)8.43975E-230.66441082811.90144462Waist circumference (cm)1.48632E-090.4214760645.115619157Waist-to-hip ratio3.93101E-100.4378102844.190208932HbA1c (%)1.48632E-090.4213469163.50851798Triglyceride1.34159E-070.371165783.494357987MQ H43, Age at hypertension diagnosis (years)0.0335602550.4523326573.38388454125(oh) vitamin d (nmol/l)3.86625E-060.3332943653.182882678ALT (U/L)3.86625E-060.3304299653.139858886BMI (kg/m2)5.2641E-050.2894060282.712077029MQ H36, Age at diabetes diagnosis (years)0.0360573380.35522.630884494Total cholesterol (mmol/l)2.94641E-050.300709222.604940422C-peptide0.0008646670.2474292152.432909903AST (u/l)0.0036832520.2137411352.333442856Body weight (kg)0.0003769370.2586879432.288926003Homocysteine5.19807E-060.3288010172.232646959Alkaline phosphatase (U/L)0.0002355150.2668354662.221953633LDL cholesterol0.0011065620.2386949182.114867708Glucose0.001583040.2306515962.060478996 Table 2Eighteen variables retained after three-criterion filtering of univariate screening results. ES: Cliff''s Delta effect size (positive values indicate higher values in CC). OR: crude odds ratio from Fisher''s exact test after median-split dichotomisation (see Methods). Variables were retained if BH-adjusted p ≤ 0.05, |Cliff''s δ|≥ 0.11, and crude OR > 2.0. Variables are ordered by descending crude OR. Abbreviations as in Table 1FieldBH-adj pCliff''s δ|ORAge (years)8.43975E-230.66441082811.90144462Waist circumference (cm)1.48632E-090.4214760645.115619157Waist-to-hip ratio3.93101E-100.4378102844.190208932HbA1c (%)1.48632E-090.4213469163.50851798Triglyceride1.34159E-070.371165783.494357987MQ H43, Age at hypertension diagnosis (years)0.0335602550.4523326573.38388454125(oh) vitamin d (nmol/l)3.86625E-060.3332943653.182882678ALT (U/L)3.86625E-060.3304299653.139858886BMI (kg/m2)5.2641E-050.2894060282.712077029MQ H36, Age at diabetes diagnosis (years)0.0360573380.35522.630884494Total cholesterol (mmol/l)2.94641E-050.300709222.604940422C-peptide0.0008646670.2474292152.432909903AST (u/l)0.0036832520.2137411352.333442856Body weight (kg)0.0003769370.2586879432.288926003Homocysteine5.19807E-060.3288010172.232646959Alkaline phosphatase (U/L)0.0002355150.2668354662.221953633LDL cholesterol0.0011065620.2386949182.114867708Glucose0.001583040.2306515962.060478996 Eighteen variables retained after three-criterion filtering of univariate screening results. ES: Cliff''s Delta effect size (positive values indicate higher values in CC). OR: crude odds ratio from Fisher''s exact test after median-split dichotomisation (see Methods). Variables were retained if BH-adjusted p ≤ 0.05, |Cliff''s δ|≥ 0.11, and crude OR > 2.0. Variables are ordered by descending crude OR. Abbreviations as in Table 1 Age-Adjusted Associations Figure 3 presents age-adjusted odds ratios for all filtered variables. Age itself remains the strongest correlate of breast cancer status (OR = 1.10 per year, 95% CI: 1.08–1.13, FDR p < 0.001), confirming that for each additional year of age, the odds of belonging to the CC group increase by approximately 10% (Table 3). Among the variables that retained independent significance after age adjustment (FDR-adjusted p ≤ 0.05), liver enzymes emerged as notable correlates. ALT (OR = 1.04, 95% CI: 1.02–1.06, FDR p = 0.006) and AST (OR = 1.05, 95% CI: 1.02–1.08, FDR p = 0.009) both showed approximately 16.5% change in log(OR) after adjustment, suggesting that while age does influence these associations, liver enzyme elevations remained independently associated with breast cancer status beyond age. Similarly, lipid parameters maintained significance, with total cholesterol (OR = 1.45, 95% CI: 1.13–1.89, FDR p = 0.018) showing a 33.8% change and LDL cholesterol (OR = 1.47, 95% CI: 1.12–1.93, FDR p = 0.018) demonstrating a 23.2% change. Despite moderate confounding by age, dyslipidaemia remains an independent discriminator between study groups.Fig. 3Forest plot of age-adjusted odds ratios (95% confidence intervals) for the 18 filtered phenotypic variables. Odds ratios represent the change in odds of breast cancer status per one-unit increase in each correlate, adjusted for age (years). Filled teal points indicate variables with FDR-adjusted p ≤ 0.05; open grey points indicate non-significant associations. The dashed vertical line marks OR = 1.0 (null). The upper confidence bound for waist-to-hip ratio is truncated at 6.0 (actual upper bound: 494.1; arrow indicates truncation). Variables are ordered by FDR-adjusted p-valueTable 3Age-adjusted logistic regression results and confounding assessment for the 18 filtered variables. For each variable, a crude model (predictor only) and an age-adjusted model (predictor + age in years) were fitted. Odds Ratio: age-adjusted OR per one-unit increase in the predictor. CI: 95% confidence interval for the age-adjusted OR. Percent change: absolute percentage change in the log-transformed odds ratio between crude and age-adjusted models, calculated as |log(OR ~ crude ~) − log(OR ~ adj ~)|/|log(OR ~ crude ~)|× 100; values ≥ 10% indicate meaningful confounding by age. FDR-adjusted p-value: Benjamini–Hochberg-corrected p-value from the age-adjusted model. Variables are ordered by ascending FDR-adjusted p-value. The waist-to-hip ratio model exhibited quasi-complete separation (see Methods); its point estimate should be interpreted with cautionFieldOdds ratioCIPercent changeFDR-adjusted pvalueAge (years)1.1031.081–1.1270.0000.000ALT (U/L)1.0381.017–1.06216.5780.006AST (U/L)1.0461.019–1.07916.4770.009Total Cholesterol (mmol/l)1.4511.129–1.88833.8420.018LDL Cholesterol (mmol/l)1.4681.123–1.93123.2490.018Homocysteine (µmol/L)1.0781.016–1.16739.9600.092Waist-to-hip ratio23.3281.160–494.05865.9950.105Triglyceride1.3590.942–1.94161.5830.211MQ H43, Age at hypertension diagnosis (years)1.0640.981–1.17316.0500.286C_peptide1.1250.956–1.31442.0820.286Glucose0.9340.839–1.024165.8440.28625(oh) vitamin d (nmol/l)1.0120.990–1.03257.0720.391Alkaline phosphatase (U/L)1.0060.994–1.01973.8840.440BMI (kg/m2)0.9820.940–1.024126.4540.511HBA1C (%)0.9310.759–1.126120.8590.566MQ H36, Age at diabetes diagnosis (years)1.0200.964–1.08460.3580.573Waist circumference (cm)1.0040.984–1.02491.8930.750Body weight (kg)0.9970.981–1.014110.5140.764* MQ_H43 and MQ_H36 (self-reported age at hypertension and diabetes diagnosis) are recorded only for participants reporting the respective condition; these models are therefore fitted on the corresponding conditional subsample and are non-significant after age adjustment. Maximum-likelihood estimate affected by quasi-complete separation. Firth penalised-likelihood refit: OR = 22.49 per unit (95% CI 1.15–464.21); the per-unit interval remains uninformatively wide owing to the narrow support of waist-to-hip ratio. Re-expressed per 0.1-unit, OR = 1.37 (95% CI 1.01–1.85). Non-significant after FDR correction Forest plot of age-adjusted odds ratios (95% confidence intervals) for the 18 filtered phenotypic variables. Odds ratios represent the change in odds of breast cancer status per one-unit increase in each correlate, adjusted for age (years). Filled teal points indicate variables with FDR-adjusted p ≤ 0.05; open grey points indicate non-significant associations. The dashed vertical line marks OR = 1.0 (null). The upper confidence bound for waist-to-hip ratio is truncated at 6.0 (actual upper bound: 494.1; arrow indicates truncation). Variables are ordered by FDR-adjusted p-value Table 3Age-adjusted logistic regression results and confounding assessment for the 18 filtered variables. For each variable, a crude model (predictor only) and an age-adjusted model (predictor + age in years) were fitted. Odds Ratio: age-adjusted OR per one-unit increase in the predictor. CI: 95% confidence interval for the age-adjusted OR. Percent change: absolute percentage change in the log-transformed odds ratio between crude and age-adjusted models, calculated as |log(OR ~ crude ~) − log(OR ~ adj ~)|/|log(OR ~ crude ~)|× 100; values ≥ 10% indicate meaningful confounding by age. FDR-adjusted p-value: Benjamini–Hochberg-corrected p-value from the age-adjusted model. Variables are ordered by ascending FDR-adjusted p-value. The waist-to-hip ratio model exhibited quasi-complete separation (see Methods); its point estimate should be interpreted with cautionFieldOdds ratioCIPercent changeFDR-adjusted pvalueAge (years)1.1031.081–1.1270.0000.000ALT (U/L)1.0381.017–1.06216.5780.006AST (U/L)1.0461.019–1.07916.4770.009Total Cholesterol (mmol/l)1.4511.129–1.88833.8420.018LDL Cholesterol (mmol/l)1.4681.123–1.93123.2490.018Homocysteine (µmol/L)1.0781.016–1.16739.9600.092Waist-to-hip ratio23.3281.160–494.05865.9950.105Triglyceride1.3590.942–1.94161.5830.211MQ H43, Age at hypertension diagnosis (years)1.0640.981–1.17316.0500.286C_peptide1.1250.956–1.31442.0820.286Glucose0.9340.839–1.024165.8440.28625(oh) vitamin d (nmol/l)1.0120.990–1.03257.0720.391Alkaline phosphatase (U/L)1.0060.994–1.01973.8840.440BMI (kg/m2)0.9820.940–1.024126.4540.511HBA1C (%)0.9310.759–1.126120.8590.566MQ H36, Age at diabetes diagnosis (years)1.0200.964–1.08460.3580.573Waist circumference (cm)1.0040.984–1.02491.8930.750Body weight (kg)0.9970.981–1.014110.5140.764* MQ_H43 and MQ_H36 (self-reported age at hypertension and diabetes diagnosis) are recorded only for participants reporting the respective condition; these models are therefore fitted on the corresponding conditional subsample and are non-significant after age adjustment. Maximum-likelihood estimate affected by quasi-complete separation. Firth penalised-likelihood refit: OR = 22.49 per unit (95% CI 1.15–464.21); the per-unit interval remains uninformatively wide owing to the narrow support of waist-to-hip ratio. Re-expressed per 0.1-unit, OR = 1.37 (95% CI 1.01–1.85). Non-significant after FDR correction Age-adjusted logistic regression results and confounding assessment for the 18 filtered variables. For each variable, a crude model (predictor only) and an age-adjusted model (predictor + age in years) were fitted. Odds Ratio: age-adjusted OR per one-unit increase in the predictor. CI: 95% confidence interval for the age-adjusted OR. Percent change: absolute percentage change in the log-transformed odds ratio between crude and age-adjusted models, calculated as |log(OR ~ crude ~) − log(OR ~ adj ~)|/|log(OR ~ crude ~)|× 100; values ≥ 10% indicate meaningful confounding by age. FDR-adjusted p-value: Benjamini–Hochberg-corrected p-value from the age-adjusted model. Variables are ordered by ascending FDR-adjusted p-value. The waist-to-hip ratio model exhibited quasi-complete separation (see Methods); its point estimate should be interpreted with caution * MQ_H43 and MQ_H36 (self-reported age at hypertension and diabetes diagnosis) are recorded only for participants reporting the respective condition; these models are therefore fitted on the corresponding conditional subsample and are non-significant after age adjustment. Maximum-likelihood estimate affected by quasi-complete separation. Firth penalised-likelihood refit: OR = 22.49 per unit (95% CI 1.15–464.21); the per-unit interval remains uninformatively wide owing to the narrow support of waist-to-hip ratio. Re-expressed per 0.1-unit, OR = 1.37 (95% CI 1.01–1.85). Non-significant after FDR correction Strikingly, several variables that appeared highly significant in the crude analysis lost their association after controlling for age, indicating substantial confounding. Metabolic markers showed near-complete confounding, with HbA1c (pct_change = 120.9%, FDR p = 0.566) and glucose (pct_change = 165.8%, FDR p = 0.286) both reversing their direction of effect, suggesting that crude associations were almost entirely driven by age. Triglycerides (pct_change = 61.6%, FDR p = 0.211) also lost significance despite retaining its direction of effect. Anthropometric measures demonstrated similarly dramatic confounding: BMI (pct_change = 126.5%, FDR p = 0.511) reversed its effect direction, waist circumference (pct_change = 91.9%, FDR p = 0.750) became non-significant, and waist-to-hip ratio (pct_change = 66.0%, FDR p = 0.105) did not retain significance after age adjustment. The age-adjusted model for this variable exhibited signs of quasi-complete separation, yielding an implausibly large point estimate (OR = 23.33) with a confidence interval spanning more than two orders of magnitude (95% CI: 1.16–494.06). This instability likely reflects the narrow distributional range of waist-to-hip ratio values (typically 0.7–1.0) combined with the modest number of cases (n = 96), which together provide insufficient information for stable maximum likelihood estimation. Penalised-likelihood refitting (Firth) returned a closely similar point estimate with an interval still spanning more than two orders of magnitude (OR = 22.49 per unit, 95% CI 1.15–464.21), confirming that the instability reflects the narrow native-scale support of waist-to-hip ratio rather than separation alone. Re-expressed per 0.1-unit increment, the Firth estimate is modest (OR = 1.37, 95% CI 1.01–1.85) and does not remain statistically significant after FDR correction (FDR p = 0.105). Waist-to-hip ratio is therefore not an independent correlate of breast cancer status after age adjustment. Other confounded variables included vitamin D (pct_change = 57.1%, FDR p = 0.391), C-peptide (pct_change = 42.1%, FDR p = 0.286), alkaline phosphatase (pct_change = 73.9%, FDR p = 0.440), and homocysteine (pct_change = 40.0%, FDR p = 0.092), which approached but did not reach significance. The magnitude of age-mediated confounding is summarised in Figs. 4 and 5 shows the distributional differences between CC and HC for the five variables retaining independent significance.Fig. 4Magnitude of age-mediated confounding across the 17 filtered phenotypic variables (Age excluded as the adjusting covariate). Bars represent the absolute percentage change in log-transformed odds ratios between crude and age-adjusted logistic regression models. Bar colour indicates confounding magnitude: blue, moderate (10–50%); orange, substantial (50–100%); red, effect reversal (> 100%). Black-bordered bars indicate variables retaining FDR-adjusted significance (p ≤ 0.05) after age adjustment. The dashed vertical line marks the conventional 10% confounding threshold; the dotted red line marks 100% (direction-of-effect reversal)Fig. 5Distributions of the five independently significant variables by group: Age (years), ALT (U/L), AST (U/L), total cholesterol (mmol/L), and LDL cholesterol (mmol/L). Boxplots compare HC (n = 471; blue) with CC (n = 96; red). Boxes span the interquartile range with the median marked; individual observations are overlaid as jittered points. Each panel is annotated with the age-adjusted odds ratio and FDR-corrected p-value from Table 3. Y-axis scales are independent across panels to accommodate differing measurement units Magnitude of age-mediated confounding across the 17 filtered phenotypic variables (Age excluded as the adjusting covariate). Bars represent the absolute percentage change in log-transformed odds ratios between crude and age-adjusted logistic regression models. Bar colour indicates confounding magnitude: blue, moderate (10–50%); orange, substantial (50–100%); red, effect reversal (> 100%). Black-bordered bars indicate variables retaining FDR-adjusted significance (p ≤ 0.05) after age adjustment. The dashed vertical line marks the conventional 10% confounding threshold; the dotted red line marks 100% (direction-of-effect reversal) Distributions of the five independently significant variables by group: Age (years), ALT (U/L), AST (U/L), total cholesterol (mmol/L), and LDL cholesterol (mmol/L). Boxplots compare HC (n = 471; blue) with CC (n = 96; red). Boxes span the interquartile range with the median marked; individual observations are overlaid as jittered points. Each panel is annotated with the age-adjusted odds ratio and FDR-corrected p-value from Table 3. Y-axis scales are independent across panels to accommodate differing measurement units These findings carry several important implications. The substantial confounding observed, with many variables showing greater than 50% change in log-OR, indicates that age differences between groups explain much of the metabolic and anthropometric variation initially observed. However, ALT, AST, total cholesterol, and LDL cholesterol remained associated with case status after conditioning on age. This indicates that the age gap alone does not account for these differences. It does not establish that they arise from a mechanism operating independently of age, and hepatic steatosis or dyslipidaemia are offered here only as candidate explanations to be tested elsewhere. Conversely, the loss of significance for HbA1c, glucose, triglycerides, and anthropometric measures after age adjustment suggests these were secondary to the age disparity rather than independent correlates of breast cancer status, in other words, metabolic syndrome features appear to be age-mediated in this cohort. Removing the Cliff''s δ and odds-ratio filters did not alter these conclusions. Among the 34 continuous variables passing FDR correction and modelled without additional filtering, only ALT, AST, total cholesterol, and LDL cholesterol retained significance after age adjustment and within-panel BH correction (Supplementary Table 4) while the large univariate hormonal differences (estradiol, total testosterone) attenuated to non-significance, consistent with their reflecting the older case age distribution. These four associations should be read as the strongest hypotheses emerging from an exploratory screen rather than as confirmed independent correlates: with 96 cases the study was powered only for medium-to-large standardised effects (Minimum detectable effect size), and BH correction controls the expected false-discovery proportion across the screen without establishing any individual association as confirmatory. Replication in an independent, adequately powered, preferably incident-case cohort is required. Age-restricted Sensitivity Analysis Models were refit on (i) the common-support age range (19–76 y; 96 cases, 460 controls) and (ii) a balanced 40–65-year window (74 cases, 144 controls; case:control ratio 1:1.9 vs 1:4.9 in the full sample). Because the case age range lay almost entirely within that of controls (Supplementary Fig. 1), common-support restriction retained all cases and 98% of controls; here ALT, AST, total cholesterol, and LDL cholesterol retained FDR significance with point estimates unchanged from the primary analysis (ALT OR 1.04, 95% CI 1.01–1.06; AST 1.04, 1.02–1.07; total cholesterol 1.42, 1.10–1.83; LDL 1.45, 1.10–1.90; all FDR p ≤ 0.032). In the narrower, age-balanced window all four estimates remained concordant in direction and magnitude but no longer reached FDR significance (FDR p 0.17–0.19), consistent with the reduced power of the smaller sample (minimum detectable effect size). No association reversed direction under either restriction. Homocysteine, borderline in the full sample, reached significance only in the 40–65 window (OR 1.21, 95% CI 1.08–1.36, FDR p = 0.019) and is noted as hypothesis-generating. These analyses indicate the liver-enzyme and lipid associations are not artefacts of the age gap, while reaffirming that the cohort is powered only for medium-to-large standardised effects. Full results are reported in Supplementary Table 3. Neither adjustment nor restriction eliminates confounding by age. Both rely on age being measured without error and modelled correctly, and neither addresses unmeasured characteristics that co-vary with age in this cohort—menopausal status, duration of metabolic comorbidity, and cumulative treatment exposure among them. Associations that remain statistically significant after these procedures are therefore associations conditional on age, not age-independent biological effects. The table for this result can be found in Supplementary Table 3. Clinical Reference-range Sensitivity Analysis Dichotomisation at clinical reference ranges corroborated the two strongest age-independent associations: abnormal ALT (≥ 33 U/L) and elevated total cholesterol (≥ 5.2 mmol/L) remained significantly associated with breast cancer status after age adjustment (ALT: OR 3.28, 95% CI 1.51–7.12; total cholesterol: OR 2.30, 95% CI 1.39–3.82; both BH p ≤ 0.006). Abnormal AST (≥ 32 U/L) was concordant in direction but did not reach significance (OR 3.05, 95% CI 0.99–9.37; BH p = 0.068), reflecting the small number of participants with abnormal values (n = 18). LDL cholesterol at the borderline-high threshold (≥ 3.4 mmol/L) was not significantly associated with breast cancer status (OR 1.53, 95% CI 0.89–2.64; BH p = 0.127). Applying the women-specific ALT limit of 19 U/L [29] classified 169 of 567 participants (30%) as abnormal, and the association attenuated to borderline significance (OR 1.67, 95% CI 1.00–2.78, p = 0.049). The ALT association therefore persists in direction and significance across both thresholds but weakens in magnitude at the lower cut-off, where abnormality becomes near ubiquitous and loses discriminatory value. Multivariable Models To assess whether the four markers retaining age-independent significance represent distinct associations rather than a single shared metabolic pattern, we fitted a series of multivariable models (Supplementary Table 5). Neither adiposity nor glycaemic status accounted for the observed associations. Adjustment for BMI in addition to age left the liver-enzyme estimates essentially unchanged (ALT OR 1.041, 95% CI 1.018–1.064; AST OR 1.047, 1.018–1.077; both BH p ≤ 0.003), with changes in the log-odds ratio of 4.2% and 2.1% respectively. Substituting waist circumference for BMI produced concordant results (ALT OR 1.040, 1.017–1.063; AST OR 1.046, 1.017–1.076). The lipid associations were likewise unchanged by adjustment for HbA1c (total cholesterol OR 1.462, 1.133–1.886; LDL OR 1.467, 1.121–1.919) or for BMI (1.443, 1.118–1.862; 1.464, 1.119–1.915), with log-odds ratio changes of 0.6–1.6%. Variance inflation factors in all these models were below 1.1, indicating no material collinearity between marker and covariate. The hepatic and lipid signals were independent of one another. In models containing age, one liver enzyme, and one lipid marker, both terms retained significance in all four pairwise combinations (ALT 1.038 with total cholesterol 1.426; ALT 1.038 with LDL 1.429; AST 1.050 with total cholesterol 1.500; AST 1.050 with LDL 1.520; all BH p ≤ 0.011), with variance inflation factors of 1.01–1.08. Likelihood-ratio tests against the age-only model confirmed this at block level: the two liver enzymes added information beyond age and both lipid markers (χ2 = 14.03, 2 df, p = 8.97 × 10⁻4), and the two lipid markers added information beyond age and both liver enzymes (χ2 = 8.58, 2 df, p = 0.014). Jointly, the four markers improved fit over age alone (χ2 = 22.53, 4 df, p = 1.57 × 10⁻4; AIC 400.5 versus 386.0). Within each pair, however, the second marker contributed nothing. AST added no information beyond age and ALT (χ2 = 1.42, 1 df, p = 0.234), and LDL cholesterol added none beyond age and total cholesterol (χ2 = 0.06, 1 df, p = 0.812); in both cases the AIC of the larger model was higher. Consistent with this, the model entering all four markers simultaneously was uninformative: no term reached significance (ALT OR 1.020, p = 0.207; AST 1.033, p = 0.086; total cholesterol 1.397, p = 0.349; LDL 1.063, p = 0.873), and the variance inflation factors for total and LDL cholesterol were 6.58 and 6.44. Because LDL cholesterol in this dataset is a calculated rather than directly assayed value, it is partly an arithmetic function of total cholesterol, and the two cannot be separately identified in a joint model. The wide confidence intervals in that model reflect this loss of identifiability rather than an absence of association. Taken together, these analyses indicate that the four markers represent two mutually independent associations with breast cancer status rather than four. Neither is explained by adiposity or by glycaemic status. Within each pair the markers are interchangeable proxies, and no basis exists in these data for preferring ALT over AST or total over LDL cholesterol. Model discrimination was quantified for the same set of models (Supplementary Table 6). Age alone discriminated breast cancer status with a c-statistic of 0.832 (95% CI 0.785–0.874). Adding each marker individually to age produced statistically detectable but numerically negligible improvement: ΔAUC 0.009 (95% CI 0.001–0.023) for ALT, 0.008 (− 0.001 to 0.025) for AST, 0.006 (− 0.001 to 0.020) for total cholesterol, and 0.008 (− 0.0003 to 0.023) for LDL cholesterol, with corresponding likelihood-ratio p-values of 3.9 × 10⁻4, 8.1 × 10⁻4, 0.004, and 0.005. Entering all four markers together with age raised the c-statistic to 0.850 (0.809–0.893), ΔAUC 0.019 (0.006–0.042). The contrast between the likelihood-ratio tests, which are significant throughout, and the bootstrap intervals for ΔAUC, three of which include zero, is expected rather than contradictory: the c-statistic is a rank-based summary whose sensitivity to an added predictor declines as the baseline model strengthens, and a baseline containing age is already strong in this cohort [31]. These results should be read as indicating that the four markers carry information about case status that age does not, while adding almost nothing to the separation of cases from controls at the individual level. No predictive, screening, or risk-stratification utility is claimed or supported by these data. Discussion This exploratory study aimed to identify phenotypic factors associated with breast cancer status in a Middle Eastern population by screening 93 demographic, anthropometric, biochemical, lifestyle, and medical history variables among Qatar Biobank participants. We compared 96 CC with 471 HC using univariate analyses followed by age-adjusted logistic regression models to identify independent associations. Of the 93 variables screened, five (including age itself) showed statistically significant differences between groups in age-adjusted analyses. Our confounding analysis indicated that ALT, AST, total cholesterol and calculated LDL remained associated with breast cancer status after adjustment for age. Multivariable modelling subsequently showed that these four markers do not constitute four distinct associations. The hepatic and lipid blocks are independent of one another, but within each block the second marker adds no information beyond the first (Supplementary Table 5). The result is therefore two associations, hepatic and lipid, neither explained by adiposity or glycaemic status. Age emerged as the strongest independent correlate of breast cancer status, with each additional year associated with a 10% increase in odds of breast cancer status. This aligns with global epidemiological data establishing advancing age as a primary risk factor, driven by cumulative genetic mutations, hormonal changes, and prolonged exposure to environmental carcinogens [5, 6]. However, our cohort reflects regional trends in the Middle East and North Africa (MENA), where breast cancer often presents at younger ages (median 45–51 years) compared to Western populations (typically > 60 years), potentially due to reproductive patterns, rapid lifestyle westernisation, and genetic factors [15, 16]. The westernized lifestyle transition underway across the Gulf provides plausible biological mediators of breast cancer risk. For example, declining parity and shortened breastfeeding are established risk-increasing exposures where each birth and each 12 months of breastfeeding independently reduce breast cancer risk by approximately 7.0% and 4.3% respectively [32], mechanistically via prolonged exposure of incompletely differentiated breast epithelium to ovulatory hormonal cycling. Another example is the rising adiposity due to western fast food diet, which is highly prevalent in the Gulf, promotes breast carcinogenesis in postmenopausal women, as adipose tissue becomes the dominant site of estrogen biosynthesis where obesity-associated inflammatory mediators (TNF-α, PGE₂, IL-6) upregulate aromatase, increasing local and circulating estrogen and driving estrogen-receptor–dependent proliferation, with concurrent hyperinsulinaemia providing an additional mitogenic stimulus [33–35]. The marked age disparity between CC (predominantly postmenopausal) and HC underscores the need for age-stratified analyses in future studies to disentangle age-related metabolic shifts from true independent associations. Critically, while age is non-modifiable, its prominence reinforces the urgency of targeted screening programs in aging MENA populations, where breast cancer incidence is projected to double by 2030 due to demographic shifts [36]. More broadly, the dominance of age in these data underscores its importance as a covariate in future risk-prediction work for this region, a goal that will require prospective, incident-case designs rather than the cross-sectional contrasts reported here. Indeed, metabolic and inflammatory biomarker associations with breast cancer have been shown to differ by menopausal status and BMI [37], underscoring the need for stratified analyses that the present sample size did not permit. Our study supports these observations, particularly through the modest elevated prevalence of central adiposity in the CC group, factors that are associated with increased postmenopausal breast cancer risk due to their impact on estrogen metabolism and chronic inflammation [7]. Similarly, the self-reported ages at diabetes diagnosis (MQ_H36) and hypertension diagnosis (MQ_H43) were higher in CC than HC, but both associations were fully attenuated after age adjustment (FDR p = 0.573 and 0.286, respectively), consistent with these variables serving as proxies for participant age rather than independent correlates of breast cancer status. This reinforces the broader pattern whereby age-correlated comorbidity measures in this cohort reflect the age disparity between groups rather than distinct biological pathways. The hepatic association, indexed interchangeably by ALT or AST, is consistent with emerging evidence linking liver enzymes to cancer-related metabolic profiles. ALT and AST have been correlated with mammographic breast density, itself a well-established risk factor [38], and may reflect subclinical metabolic perturbations relevant to breast cancer biology. These associations may reflect subclinical hepatic steatosis or insulin resistance, conditions prevalent in populations with high rates of central adiposity and metabolic syndrome, as observed in the Gulf region. This interpretation is constrained, however, by the multivariable results: adjustment for BMI, waist circumference or HbA1c left the ALT and AST estimates essentially unchanged (Supplementary Table 5), so the hepatic signal is not a proxy for measured adiposity or glycaemia in this cohort. Whether it reflects an unmeasured hepatic process cannot be tested here, as liver imaging and steatosis indices were not available in the QBB extract. However, a critical interpretive caveat applies. Because QBB phenotype data were collected at enrolment, and CC were identified through pre-existing EMR diagnoses, some participants in the CC group may have already received systemic therapy at the time of biospecimen collection. Systemic therapies commonly used in breast cancer management include chemotherapy, endocrine therapy, and targeted agents, all of which are known to induce hepatocellular injury manifesting as ALT and AST elevation [39], [40]. Consequently, the observed liver enzyme differences between CC and HC could reflect treatment-induced hepatotoxicity or a survivorship-related metabolic state rather than a pre-diagnostic phenotype. Without data on treatment timing relative to QBB enrolment, this distinction cannot be resolved in the present study, and the ALT/AST findings should be interpreted with this ambiguity in mind. The relationship between circulating lipids and breast cancer risk appears complex and partly inconsistent. Large meta-analyses of prospective cohorts involving over one million participants have generally found little or no association between total or LDL cholesterol and breast cancer incidence [41–43]. In contrast, Mendelian randomization studies suggest that genetically elevated LDL cholesterol is associated with higher breast cancer risk, particularly for estrogen receptor–positive disease [44, 45]. More recent observational data, including large Asian cohorts, also report modestly increased breast cancer risk at higher LDL cholesterol levels [46], supporting a potential contributory role of dyslipidaemia despite overall modest effect sizes. In the present data, however, total and LDL cholesterol cannot be treated as separate findings. LDL was derived by the Friedewald equation and is therefore partly an arithmetic function of total cholesterol,the two are not separately identifiable when entered jointly (Supplementary Table 5). The lipid association reported here is a single signal, and these data provide no basis for attributing it specifically to the LDL fraction. Limitations The limitations of our study are directly associated with the inherent characteristics of the QBB data. While the QBB provides a uniform and well-characterised cohort, several important constraints persist that affect the interpretation of our findings. The cross-sectional design of this study prevents determination of temporal relationships between associated factors and breast cancer diagnosis. We considered restricting analysis to incident cases diagnosed after enrolment, or a nested case–control design, as suggested. However, the QBB extract contains no post-enrolment follow-up and our IRB approval does not permit re-contact. A prospective or nested redesign is therefore not feasible within the current data-access and ethics framework. Accordingly, all associations reported describe cross-sectional differences by breast cancer status and should not be interpreted as risk factors or predictors of incidence. The approximately 20-year median age difference between CC and HC remains the dominant structural feature of these data. Statistical adjustment for age and restriction to overlapping age ranges reduce but cannot remove this confounding, and residual confounding by age and by age-correlated characteristics not captured in the QBB extract is likely. No age-adjusted association reported here should be read as evidence of an independent biological mechanism. Furthermore, the modest sample size of CC (n = 96) may have limited statistical power to detect independent associations for variables other than age. Several age-adjusted logistic regression models exhibited signs of quasi-complete separation, most notably for waist-to-hip ratio, where the narrow predictor range relative to the small number of cases produced an unstable point estimate (OR = 23.33, 95% CI: 1.16–494.06). Standard maximum likelihood estimation is known to be unreliable under these conditions, and the resulting odds ratios and confidence intervals should be interpreted with caution [27]. We therefore refitted the affected waist-to-hip ratio model using Firth''s penalised-likelihood logistic regression, which returned a finite but still imprecise estimate (per 0.1-unit OR = 1.37, 95% CI 1.01–1.85,non-significant after FDR correction), confirming that waist-to-hip ratio is not an independent correlate of breast cancer status; affected models are flagged accordingly. Another limitation of this study is the control definition: controls were required to have neither a physician-documented diagnosis of any disease nor any self-reported disease history. This produces an exceptionally healthy-volunteer, or ''super-normal'', control group. Such selection is a recognised feature of volunteer-based biobank cohorts and can distort exposure–outcome contrasts [47]. Because controls are systematically healthier than the general population, case–control differences in metabolic and biochemical markers are likely inflated in magnitude relative to a comparison against a representative cancer-free group. A less restrictive control group, defined solely by the absence of a breast cancer diagnosis rather than the absence of any disease, would provide a more representative comparison, but would not address the cross-sectional, prevalent-case structure that limits causal interpretation. The univariate screening stage dichotomised continuous variables at the pooled-sample median to compute crude odds ratios. This is statistically inefficient and carries three specific costs: loss of statistical power relative to analysis on the continuous scale; loss of information, since all within-category variation is discarded and participants either side of the cut-point are treated as maximally different while those at opposite extremes of a category are treated as identical; and dependence on an arbitrary, data-derived cut-point that has no clinical or biological meaning and would shift with the composition of the sample [23]. We mitigated but did not remove this limitation by confining the median split to the screening stage — all inferential models used predictors on their continuous scale — and by repeating the analysis for the four primary markers using established clinical reference thresholds rather than sample-derived medians (Table 4). Variables excluded at the screening stage nonetheless remain subject to the power loss that dichotomisation entails, and some may have been dropped for that reason alone.Table 4Clinical reference-range sensitivity analysis. Age-adjusted odds ratios for abnormal-versus-normal classification of the four markers retaining age-independent significance, dichotomised at established clinical thresholds, with age as a continuous covariate. ALT is reported at the traditional limit (≥ 33 U/L; primary) and the women-specific limit (≥ 19 U/L; secondary, [29]. ‘Abnormal n’ refers to the number of participants at or above the threshold. p-values are Benjamini–Hochberg-adjusted across the four primary tests (ALT-33, AST, total cholesterol, LDL)MarkersThresholdAbnormal nAge-adj OR (95% CI)Raw pBH pALT ≥ 33 U/L (primary)403.28 (1.51–7.12)0.00270.0054ALT ≥ 19 U/L (secondary)1691.67 (1.00–2.78)0.049244868 = AST ≥ 32 U/L183.05 (0.99–9.37)0.0510.068Total cholesterol ≥ 5.2 mmol/L1782.30 (1.39–3.82)0.00130.005LDL ≥ 3.4 mmol/L1181.53 (0.89–2.64)0.1270.127 Table 4Clinical reference-range sensitivity analysis. Age-adjusted odds ratios for abnormal-versus-normal classification of the four markers retaining age-independent significance, dichotomised at established clinical thresholds, with age as a continuous covariate. ALT is reported at the traditional limit (≥ 33 U/L; primary) and the women-specific limit (≥ 19 U/L; secondary, [29]. ‘Abnormal n’ refers to the number of participants at or above the threshold. p-values are Benjamini–Hochberg-adjusted across the four primary tests (ALT-33, AST, total cholesterol, LDL)MarkersThresholdAbnormal nAge-adj OR (95% CI)Raw pBH pALT ≥ 33 U/L (primary)403.28 (1.51–7.12)0.00270.0054ALT ≥ 19 U/L (secondary)1691.67 (1.00–2.78)0.049244868 = AST ≥ 32 U/L183.05 (0.99–9.37)0.0510.068Total cholesterol ≥ 5.2 mmol/L1782.30 (1.39–3.82)0.00130.005LDL ≥ 3.4 mmol/L1181.53 (0.89–2.64)0.1270.127 Clinical reference-range sensitivity analysis. Age-adjusted odds ratios for abnormal-versus-normal classification of the four markers retaining age-independent significance, dichotomised at established clinical thresholds, with age as a continuous covariate. ALT is reported at the traditional limit (≥ 33 U/L; primary) and the women-specific limit (≥ 19 U/L; secondary, [29]. ‘Abnormal n’ refers to the number of participants at or above the threshold. p-values are Benjamini–Hochberg-adjusted across the four primary tests (ALT-33, AST, total cholesterol, LDL) Two further constraints follow from the multivariable analysis. First, the four markers retaining significance after age adjustment represent two independent signals rather than four: ALT and AST are interchangeable indices of the hepatic association, and total and calculated LDL cholesterol of the lipid association (Supplementary Table 5). No inference should therefore be drawn as to which enzyme or which lipid fraction is involved, and the count of "four markers" should not be read as four pieces of evidence. Second, the incremental discrimination attributable to these markers is negligible. Adding each marker to age raised the c-statistic by 0.006–0.009, and all four together by 0.019 (Supplementary Table 6). Associations detectable at the cohort level by likelihood-ratio test therefore do not translate into separation of individual cases from controls, and nothing in these data supports a screening or risk-stratification application. The absence of clinical oncology data represents the most consequential limitation of this study. The QBB research extract does not include tumour characteristics (hormone receptor status, stage, grade) or treatment history for participants with an EMR-documented breast cancer diagnosis, nor does it confirm that each recorded diagnosis represents a primary breast malignancy rather than secondary breast involvement. Exclusion of participants with other documented malignancies makes the latter unlikely but does not exclude it. Because QBB phenotype data were collected at enrolment rather than at a defined time point relative to diagnosis, we cannot determine whether individual cases were treatment-naïve, undergoing active therapy, or in post-treatment follow-up at the time of measurement. This ambiguity is particularly consequential for interpreting the ALT and AST findings, where chemotherapy-induced hepatotoxicity represents a plausible alternative to a pre-diagnostic metabolic explanation, as discussed above. More broadly, the inability to distinguish pre-diagnostic biomarker levels from treatment- or survivorship-related changes applies across all biochemical variables examined and limits the degree to which any observed association can be attributed to breast cancer biology rather than its management. This limitation is particularly evident in our unexpected vitamin D results, where CC showed higher rather than lower levels compared to HC. Without knowledge of the tumour subtype and the supplementation pattern, such biomarker findings are difficult to interpret meaningfully. In addition, biomarker trajectories and absolute levels are known to differ by tumour subtype, stage, and systemic therapy, but such tumour-level characteristics were not available in the present dataset, precluding subtype-stratified analyses. Relatedly, menopausal status was unavailable and could not be reliably reconstructed: no menopausal field or age at menopause was recorded, gonadotrophins were not measured, and estradiol is an unreliable surrogate here owing to endocrine-therapy confounding in cases and substantial differential missingness. Direct adjustment or stratification by menopausal status was therefore not possible. Because menopausal status is overwhelmingly age-determined, the inclusion of age in all adjusted models and the age-restricted sensitivity analyses (Supplementary Table 3) absorb its major age-driven component; nevertheless, residual confounding by menopausal status independent of age cannot be excluded, and metabolic and inflammatory biomarker associations are known to differ by menopausal stage [37]. This study should be considered exploratory, generating hypotheses for future research rather than providing definitive clinical guidance for risk stratification in Middle Eastern populations. Future research priorities should include larger prospective cohort studies or registry-linked biobank analyses with detailed clinical characterization of participants with disease, including tumour stage, grade, molecular subtype (e.g., ER/PR/HER2), and treatment history. Such studies will be essential to disentangle subtype-specific biomarker patterns and to develop robust, population-specific risk prediction models. Conclusion This exploratory cross-sectional analysis of the Qatar Biobank (QBB) cohort provides preliminary insights into phenotypic associations with breast cancer status in an understudied Middle Eastern population. By screening 93 variables among 96 CC and 471 HC, we identified age as the strongest independent correlate of breast cancer status (OR = 1.10 per year, 95% CI: 1.08–1.13, p < 0.001). Liver enzymes (ALT: OR = 1.04, 95% CI: 1.02–1.06, p = 0.006; AST: OR = 1.05, 95% CI: 1.02–1.08, p = 0.009) and lipid parameters (total cholesterol: OR = 1.45, 95% CI: 1.13–1.89, p = 0.018; LDL cholesterol: OR = 1.47, 95% CI: 1.12–1.93, p = 0.018) retained independent significance after age adjustment, while metabolic and anthropometric markers (e.g., HbA1c, BMI, waist circumference) showed substantial age-mediated attenuation (> 50% change in log-OR). Multivariable modelling indicates these four markers represent two mutually independent associations i.e. hepatic and lipid, rather than four distinct signals, and that neither is accounted for by adiposity or glycaemic status (Supplementary Table 5). These findings are subject to constraints that preclude etiologic interpretation. The cross-sectional design with prevalent cases means the temporal ordering of biomarker level and diagnosis is unknown. In particular, the elevated liver enzymes are compatible with chemotherapy-induced hepatotoxicity or a survivorship-related metabolic state, and the elevated vitamin D observed in cases plausibly reflects post-diagnosis supplementation. The modest case count (n = 96), the absence of tumour characteristics and treatment history, the healthy-volunteer control definition, and residual confounding by age each further limit interpretation. Extending QBB infrastructure to support longitudinal follow-up and linkage to tumour characteristics and treatment records would remove several of these constraints for future analyses in this cohort. We therefore draw no causal or predictive conclusion. The four markers retaining significance after age adjustment are the strongest hypotheses generated by this screen and nothing more. Testing them requires an independent, adequately powered cohort with incident cases, pre-diagnostic sampling, and linked tumour and treatment data. Supplementary Information Below is the link to the electronic supplementary material.Supplementary file1 (XLSX 35 KB)Supplementary file2 (XLSX 24 KB)Supplementary file3 (XLSX 11 KB)Supplementary file4 (XLSX 14 KB)Supplementary file5 (XLSX 15 KB)Supplementary file6 (XLSX 10 KB)Supplementary file7 (PDF 15 KB) Supplementary file1 (XLSX 35 KB) Supplementary file2 (XLSX 24 KB) Supplementary file3 (XLSX 11 KB) Supplementary file4 (XLSX 14 KB) Supplementary file5 (XLSX 15 KB) Supplementary file6 (XLSX 10 KB) Supplementary file7 (PDF 15 KB)