Calibrated dietary patterns and cancer risk in the women’s health initiative cohorts

Abstract

We developed calibration equations using metabolomics from fasting blood and 24-hour urine for Healthy Eating Index 2010 (HEI-2010) and Alternative Healthy Eating Index 2010 to address measurement error from self-reported diet. We examined associations between metabolomic-calibrated dietary patterns and cancer risk in the Women's Health Initiative (WHI) (n = 108 522). Metabolomic signatures were created from a WHI Feeding (n = 153; 2010-2014) and WHI Observational Study (n = 450; 2006-2009). Dietary patterns were regressed on metabolites using the feeding study food intake records. Metabolomic-based dietary patterns were estimated from 24-hour dietary recalls, food frequency questionnaire and 4-day food record in the Observational Study using a stepwise approach. Cox regression estimated cancer risk of metabolomic-calibrated dietary patterns with a median follow-up of 15.8 years. Adjusted R² for HEI-2010 and Alternative Healthy Eating Index 2010 calibration equations were 57.5% and 48.8% for food frequency questionnaire, 61.6% and 62.6% for 4-day food record, and 52.5% and 53.2% for dietary recalls. Without calibration, a 20% increment in HEI-2010 was associated with lower risk of colorectal (HR, 0.94; 95% CI, 0.90-0.99), lung (HR, 0.90; 95% CI, 0.86-0.94), bladder (HR, 0.86; 95% CI, 0.75-0.99), and total invasive cancers (HR, 0.98; 95% CI, 0.96-0.99). With metabolomic calibration, higher HEI-2010 was associated with lower risk of lung (HR, 0.79; 95% CI, 0.71-0.88) and total invasive cancers (HR, 0.96; 95% CI, 0.92-1.00). Metabolomic-calibrated dietary patterns might mitigate measurement errors and strengthen diet–cancer associations.

Trial registration:  www.whi.org

Introduction

Dietary guidelines now emphasize dietary patterns over individual nutrients, as dietary patterns reflect long-term exposure to the whole diet, which may offer the most suitable approach to assessing diet-related cancer risk.^1-3 Research has demonstrated that adherence to various dietary patterns, such as Healthy Eating Index (HEI), Alternative Healthy Eating Index (AHEI), Mediterranean diet score, Dietary Approaches to Stop Hypertension, is associated with lower risk of colorectal, breast, prostate, and pancreatic cancers.^4-6 These indices often rely on self-reported dietary intake from food frequency questionnaires (FFQs) that are prone to measurement error and that bias dietary pattern and disease risk estimation.^7-9 Objective dietary biomarkers can be useful to evaluate the reliability of self-reported dietary patterns and improve the reliability of diet–disease estimations.

Metabolomics measures small molecules in bodily fluids, such as blood and urine, including both endogenously produced compounds and compounds from exogenous exposure (ie, diet).¹⁰ Using mass spectrometry and high-dimensional bioinformatics, metabolomics detects thousands of molecules that may reflect ingested and metabolized foods and nutrients. Recently, metabolomics has provided unique opportunities to identify dietary patterns using objectively measured metabolites from biospecimens (eg, blood, urine).¹¹ Several studies have incorporated serum and urine metabolomics with self-reported data to assess dietary patterns; however, most examined the cross-sectional correlations between metabolomics and self-reported dietary patterns,^12-15 which precludes making inferences about the direction of associations and investigation of time-related associations. The cross-sectional approach is also highly dependent on observational data (eg, Framingham Heart Study, Cancer Prevention Study II Nutrition Cohort) that may result in spurious associations; replication in separate datasets and across study designs is needed.¹² An alternative approach was developed from previous studies to identify nutritional biomarkers (such as urinary nitrogen, sodium, serum tocopherols, carotenoids, retinol, and phospholipid fatty acids) using biospecimens collected in a controlled feeding study (types and quantities of specific foods and beverages that are known) and applying the identified biomarkers to develop regression calibration models to correct measurement error in the self-reported dietary data.^3,16-19 This approach has not yet been used for calibrating the self-report of dietary patterns using serum and urine metabolomic profiles.

Biomarker-calibrated dietary patterns allow objective dietary measures to provide empirical evidence and link dietary patterns with risk of diet-related chronic diseases, such as cancer. Using common nutritional biomarkers, we have previously demonstrated that higher biomarker-calibrated HEI-2010 scores were associated with lower risk of colorectal and breast cancer and lower cancer mortality.²⁰ In the same study, we observed that body mass index (BMI) played a role in the diet-disease associations that deserves further exploration as less is known about the relationship between nutritional biomarkers and dietary patterns by BMI subgroups. It is possible that due to the metabolic variations by BMI,²¹ the association between calibrated dietary patterns and cancer risk differ by BMI. The current study aimed to (1) identify serum and urine metabolomic biomarkers for the development of calibration equations for dietary patterns (HEI-2010, AHEI-2010); (2) apply the metabolomic-calibrated dietary patterns to determine the association between calibrated dietary patterns and cancer risk; and (3) examine whether the association between metabolomic-calibrated dietary patterns and cancer risk differed by BMI categories.

Methods

Study population and design

The Women's Health Initiative (WHI) is a longitudinal study dedicated to enhancing chronic disease prevention strategies in postmenopausal women. Briefly, during 1993 to 1998, postmenopausal women (aged 50-79 years) were enrolled from 40 clinical centers across the United States into three overlapping randomized clinical trials (WHI-CT) (n = 68 132) and an observational study (WHI-OS) (n = 93 676). The study details have been described previously.^22,23 As of February 2023, 50 965 women were alive and in active follow-up. All study procedures were reviewed and approved by the institutional review boards of the participating WHI sites and the Clinical Coordinating Center. All participants provided written informed consent. WHI is registered at clinicaltrials.gov as NCT00000611.

The WHI Nutrition and Physical Activity Assessment Study Observational Study (NPAAS-OS, 2006-2009) and the WHI Nutrition and Physical Activity Assessment Study Feeding Study (NPAAS-FS, 2010-2014) were ancillary studies aimed to better understand the measurement properties of both self-reported diet and physical activity assessment (Figure 1).^7,17 These protocols received institutional review board approval by the Fred Hutchinson Cancer Center and each participating clinical center, and women signed informed consent that was in addition to their on-going written consent to participate in the WHI and WHI-Extensions.

Figure 1.

Study samples and flow in the WHI cohorts and in NPAAS cohorts.

In 2006- to 2009, the NPAAS-OS enrolled 450 WHI-OS women from 9 WHI Clinical Centers to complete a doubly labeled water (DLW) protocol for assessment of total energy expenditure (TEE), a 24-hour urine collection to measure urinary nitrogen (dietary protein intake), fasting blood draws, anthropometry, and three self-reported dietary assessments (FFQ, 4-day food record [4DFR], and three 24-hour recalls). Participants were free-living, following usual diet and physical activity habits, and attended two clinic visits to complete study procedures.^7,19 The NPAAS-OS oversampled participants with younger age, higher BMI, and non-White race/ethnicity. In 2010 to 2014, the NPAAS-FS enrolled 153 women from WHI-OS, WHI Dietary Modification Trial control arm (WHI-DM-C), or the WHI Hormone Therapy Trial who lived in metropolitan Seattle area.¹⁷ Participants completed a 2-week controlled diet where they were provided all food and most beverages. Each participant's controlled diet was individualized to reflect their usual diet as initially assessed from 4DFR with adjustments (eg, to total energy) following nutritionist review. This design minimized perturbations in blood and urine measurements while preserving intake variation in the study population. The characteristics of participants from the NPAAS-FS and NPAAS-OS are presented in Table S1. Details of study design, food preparation details, and participant adherence measures are published.¹⁷ These consumed menus were entered into Nutrition Data System for Research© software (University of Minnesota) to calculate food and nutrient intakes. The current analysis included participants from the OS and DM-C if they had a valid baseline FFQ based on daily energy intake (>600 or <5000 kcals/days), returned at least one follow-up questionnaire during the study, and had no missing data for any of a common set of potential confounding variables (Figures 1 and 2). These two studies, NPAAS-FS and NPAAS-OS, objectively measured metabolomic biomarkers, including serum aqueous metabolites (analyzed by targeted LC–MS/MS via authentic commercially obtained compounds), serum lipids (analyzed by Lipidyzer), 24-hour urine metabolites (analyzed by ¹H NMR spectroscopy using a Bruker Avance III 800 MHz NMR spectrometer), and urine metabolites (analyzed by global gas chromatography-mass spectrometry (GC–MS) using an Agilent 7890/5975C GC–MS instrument). All metabolomics profiling was performed at the University of Washington's Northwest Metabolomics Research Center. Details on the metabolite profiling are provided in Appendix S1.

Figure 2.

Flow diagram of participants included in the current analyses from the WHI OS and DM-C cohorts.

Data and biospecimen collection

All participants completed WHI baseline clinic visits where protocol-driven measures of height and weight were collected by study staff, along with self-reported personal and family medical history, demographic characteristics, lifestyle habits,²⁴ and usual diet using a FFQ designed specifically for WHI.²⁵ The detailed methods for metabolite profiling and other nutritional biomarker analyses were included in Appendix S1.

Dietary assessment

The self-reported WHI FFQ included 122-line items for foods and food groups, 19 adjustment questions that add precision to estimate of fat and whole grain intake, and two summary questions. The baseline FFQs were analyzed by Nutrition Data System for Research© (University of Minnesota, version 2005) using the US Department of Agriculture’s Nutrient Database for Standard Reference plus manufacturer's information.¹⁷ For each participant, NPAAS-FS consumed diets and NPAAS-OS self-reported diets (from FFQ, 4DFR, 24-hour Recall [24-HR]) were used to calculate dietary pattern scores, including HEI-2010 and AHEI-2010. Healthy Eating Index 2010 is designed to assess adherence to the Dietary Guidelines for Americans.²⁶ It is composed of 12 components, including total fruits, whole fruits, total vegetables, greens and beans, whole grains, refined grains, sodium, empty calories, dairy, fatty acids, total protein foods, and seafood and plant proteins. Score for each component varies from 5 to 20, and the total HEI-2010 score ranges from 0 to 100, with higher score indicating better adherent with dietary guidelines.²⁶ Alternative Healthy Eating Index 2010 is slightly different from the HEI as it includes 11 components: fruits, vegetables, whole grains, sugar-sweetened beverages, nuts and legumes, red or processed meats, trans fat, long-chain fats, polyunsaturated fatty acid, sodium, and alcohol. Total AHEI-2010 score ranges from 0 to 110, with higher score indicating a healthier dietary pattern.²⁷ Healthy Eating Index 2010 and AHEI-2010 were used for the baseline dietary pattern scores as the timing of the NPAAS studies was close to when the 2010 Dietary Guidelines for Americans, HEI-2010, and AHEI-2010 were published.^26,27

Calibration of dietary patterns

Biomarker calibration of dietary patterns was a two-stage process.³ First, NPAAS-FS data and biospecimens were used to identify the metabolomic biomarkers reflecting HEI-2010 and AHEI-2010 dietary pattern scores. Then NPAAS-OS data and biospecimens were used to develop the calibration equations for each dietary pattern score. For stage 1, we used NPAAS-FS data for discovery of metabolomic signatures of dietary patterns by regressing each dietary score on all log-transformed serum metabolites, all log-transformed urine metabolites, or both (log-transformed) serum and urine metabolites, in addition to established nutritional biomarkers and participant characteristics with Least Absolute Shrinkage and Selection Operator (LASSO) penalty for variable selection¹¹ (Figure 1). Model R² ≥ 0.36 was required for an acceptable biomarker pattern signature.^3,17 Leave one out cross validation R² was computed for the final set of selected predictors to reduce overfitting. For HEI-2010 and AHEI-2010, the best models were selected based on the highest cross-validated R². The leave one out cross validation R² for HEI-2010 and AHEI-2010 were 0.35 and 0.50, and the model R² were 0.54 and 0.60, respectively (Table S2). These biomarker signatures for HEI-2010 and AHEI-2010 were brought forward for stage 2 analysis in NPAAS-OS.

For stage 2, the β coefficients from the NPAAS-FS models (listed in Table 1) were used to compute predicted metabolomic-based dietary pattern scores in NPAAS-OS using serum and urine metabolite levels as well as serum phospholipid fatty acids and nutritional biomarkers from NPAAS-OS (Figure 1). Self-reported dietary pattern scores (from FFQ, 4DFR, and 24-HR) and participant characteristics were considered for entry into the models using a stepwise approach with P < .10 for entry/removal.²⁸ To explore whether the calibration equation varied by BMI, we included a product (interaction) term between BMI and self-reported dietary pattern score (HEI-2010 or AHEI-2010). FFQ was used for the risk association analysis since only FFQ was available on all WHI participants. Details of calibration equation development for other dietary variables while using both NPAAS-OS and NPAAS-FS are published and included in Appendix S1.^8,11,16,18

Table 1.

Estimates from regression of dietary pattern scores on log metabolomic biomarker levels in the WHI NPAAS-FS (n = 153), 2010-2014.^*

Dietary pattern	β (SE)	R 2  (%)	Cross-validated R 2 a  (%)
HEI-2010
Intercept	−118.76 (61.56)
Cholesterol ester (CE 15:0)	−7.97 (5.01)	14.2	9.1
Indole-3-propionateb,c	3.44 (1.51)	12.2	7.8
Ceramide (CER 22:0)	−17.83 (7.54)	5.7	3.6
β carotene	2.43 (1.16)	8.0	5.2
Hippuric acidc	1.84 (0.96)	2.0	1.3
Uracilc	4.59 (3.12)	3.7	2.4
Uridinec	6.93 (4.69)	1.3	0.9
Diacylglycerol (DAG 18:1, 22:6)	1.38 (1.00)	3.2	2.1
Docosopentaenoic acid	−2.61 (2.09)	1.8	1.2
Trans, cis-linoelaidic acidb	−3.64 (3.59)	1.7	1.1
Triacylglycerol (TAG 51:2, FA16:0)	−3.53 (4.77)	0.4	0.2
13-Octadecenoic acid	−2.06 (2.53)	0.2	0.2
Cholesterol ester (CE 18:0)b	−2.83 (7.63)	0.0	0.0
Cholesterol ester (CE 14:1)	−0.56 (2.73)	0.0	0.0
Triacylglycerol (TAG 50:3, FA 16:0)b	−0.70 (4.69)	0.0	0.0
Total		54.4	35.1
AHEI-2010
Intercept	−51.55 (27.74)
Triacylglycerol (TAG 58:8, FA 22:6)	11.03 (5.75)	19.1	16.0
Triacylglycerol (TAG 58:9, FA22:6)	−8.52 (5.80)	1.0	0.8
Cellobiosec	−2.63 (0.44)	7.1	6.0
Indole-3-propionateb,c	4.81 (1.19)	7.9	6.6
Triacylglycerol (TAG 50:2, FA 16:0)	−9.24 (6.16)	3.1	2.6
Cholesterol ester (CE 18:0)b	−20.68 (6.72)	9.2	7.7
Bilobalide	2.12 (0.65)	3.1	2.6
Cholesterol ester (CE 22:4)	−9.98 (3.88)	4.1	3.5
Phosphatidylethanolamine (PE 18:0, 22:6)	4.64 (1.98)	2.9	2.4
Phosphatidylethanolamine (PE 18:0, 20:3)	−3.41 (1.82)	1.5	1.2
Triacylglycerol (TAG 53:3, 16:0)	−3.81 (4.45)	0.3	0.2
Cis, trans-linoelaidic acid	−2.97 (4.21)	0.1	0.1
Trans, cis-linoelaidic acidb	1.52 (3.21)	0.1	0.1
Phosphatidylcholine (PC 18:0, 22:4)	−1.07 (2.53)	0.1	0.0
Triacylglycerol (TAG 50:3, FA 16:0)b	1.00 (6.17)	0.0	0.0
Total		59.6	50.0

Abbreviations: NPAAS-FS, Nutrition and Physical Activity Assessment Study Feeding Study; WHI, Women’s Health Initiative.

^a Cross-validated R² was calculated using leave one out cross validation method.

^b Urine metabolites.

^c Common metabolite signatures identified in both HEI-2010 and AHEI-2010.

^*In FA X:A, CE X:A, CER X:A, DAG X:A, Y:B, PE X:A, Y:B, PC X:A, Y:B, X and Y indicate the number of carbon atoms and A and B indicate the number of double bonds in the fatty acid chains. In TAG X:A, X indicates the total number of carbon atoms and A indicates the total number of double bonds in the three fatty acid chains.

Outcome ascertainment

Clinical outcomes were reported biannually in the WHI-CT and annually in the WHI-OS by self-administered questionnaire throughout the time from WHI enrollment to the end of the intervention period (2005) and annually thereafter in both cohorts.²⁹ An initial report of invasive cancer during cohort follow-up was confirmed by review of medical records and pathology reports by physician-adjudicators. All invasive cancers, except nonmelanoma skin cancer, were centrally coded using the National Cancer Institute’s Surveillance, Epidemiology, and End Results procedures.²⁹ Cancer outcomes through February 2023 were included in this analysis, providing a median follow-up duration of about 15.8 years. We defined disease occurrence time for a “case” developing a cancer outcome as days from “baseline” (year 1 in WHI-DM-C and enrollment in the WHI-OS) to diagnosis. We defined censoring time for “non-cases” as days from baseline to the earliest of date of death without the outcome under study, last contact, or February 19, 2023.

Statistical analysis

For endometrial cancer analyses, women who had hysterectomy at baseline (WHI enrollment for OS participants and year 1 for DM-C participants) were excluded, as were those missing disease specific covariates generally. After exclusions, 108 522 participants were included in the analyses (Figure 2).

Descriptive statistics characterized demographic characteristics of WHI participants. Cox proportional hazards regression was used to examine the association between self-reported dietary scores (from FFQ, log transformed) with and without metabolomic calibration and incident of total invasive cancers, obesity-related cancers (breast, colon, rectum, endometrium, or kidney cancer), and by cancer type. The estimated HRs model a 20% increment in dietary pattern score in relation to each outcome where a higher score reflects a better diet quality. Participants are excluded in the Cox proportional hazards model if they reported a history of any cancer (apart from non-melanoma skin cancer) at baseline. All Cox regression models used baseline hazard rates stratified on study component (DM-C or OS), hormone therapy trial status (estrogen plus progestin, estrogen plus progestin placebo, estrogen alone, estrogen-alone placebo, not randomized), age at enrollment (50-54, 55-59, 60-64, 65-69, 70-74, ≥75 years), and BMI categories (<25, 25-30, ≥30 kg/m²), and with adjustment for a disease-specific set of potential confounding factors (including age, BMI, race/ethnicity, education, season of FFQ completion, current smoking status, supplement use, and recreational physical activity per week). For display purposes we show HRs (95% CIs) for a 20% increment in the dietary variable in Results. To account for the differences between individuals included and excluded from the study, sensitivity analysis was conducted using inverse probability weighting (IPW) of inclusion to assess cancer incidence HRs for 20% increment in the dietary pattern scores.

To investigate whether BMI modified the association between metabolomic-calibrated dietary pattern and cancer risks, we examined a product (interaction) term of BMI and metabolomic-calibrated log-transformed dietary pattern scores in the Cox model. Specifically, we estimated the HR and 95% CI between metabolomic-calibrated dietary patterns and cancer risk among normal weight or underweight (n = 36 126), overweight (n = 32 460), and obese individuals (n = 26 044). Wald tests were used to examine whether the interaction term was significant (P values < .10) in the full model.³⁰ All tests were 2-sided and considered P values < .05 to be statistically significant for the main effect.

Results

Baseline demographic and lifestyle characteristics of the participants from the WHI-OS (n = 86 291) and WHI DM-C (n = 22 231) used in the disease association models are presented in Table 2. Discovery of metabolite signatures of dietary patterns are provided in Table 1. Metabolite signatures identified for HEI-2010 include cholesterol ester (CE 15:0), indole-3-propionate, ceramide (CER 22:0), β-carotene, hippuric acid, uracil, uridine, diacylglycerol (DAG 18:1/22:6), docosopentaenoic acid, trans-cis linoelaidic acid, triacylglycerol (TAG, 51:2/FA 16:0), 13-octadecenoic acid, cholesterol ester (CE 18:0), cholesterol ester (CE 14:1), and triacylglycerol (TAG 50:3/FA 16:0). Metabolite signatures identified for AHEI-2010 include triacylglycerol (TAG 58:8/FA22:6), triacylglycerol (TAG 58:9/FA 22:6), cellobiose, indole-3-propionate, triacylglycerol (TAG 50:2/FA16:0), cholesterol ester (CE 18:0), bilobalide, cholesterol ester (CE 22:4), phosphatidylethanolamine (PE 18:0/22:6), phosphatidylethanolamine (PE 18:0/20:3), triacylglycerol (TAG 53:3/FA 16:0), cis-trans linoelaidic acid, trans-cis linoelaidic acid, phosphatidylcholine (PC 18:0/22:4), triacylglycerol (TAG 50:3/FA 16:0).

Table 2.

Baseline demographic and lifestyle characteristics of women from the WHI DM-C and OS, enrolled during 1993-1998 at 40 US clinical centers.^a

Demographics	OS (n = 86 291)		DM-C (n = 22 231)		P
	N	%	N	%
Age, years					<.001
50-54	11 404	13.2	1940	8.7
55-59	15 998	18.5	4564	20.5
60-64	18 814	21.8	5602	25.2
65-69	19 257	22.3	5161	23.2
70-74	14 356	16.6	3418	15.4
≥75	6462	7.5	1546	7.0
BMI (kg/m2)					<.001
<25	35 425	41.1	5950	26.8
25- < 30	29 335	34.0	7846	35.3
≥30	21 531	25.0	8435	37.9
Race					<.001
American Indian/Alaskan Native	277	0.3	56	0.3
Asian	2373	2.7	542	2.4
Native Hawaiian or other Pacific Islander	54	0.1	14	0.1
Black	6280	7.3	2058	9.3
White	74 710	86.6	18 986	85.4
More than one race	873	1.0	311	1.4
Unknown or not reported	1724	2.0	264	1.2
Ethnicity					<.001
Not Hispanic/Latino	82 012	95.0	21 253	95.6
Hispanic/Latino	3432	4.0	869	3.9
Unknown or not reported	847	1.0	109	0.5
Education					<.001
High school or less	4086	4.7	946	4.3
High school diploma/GED	13 899	16.1	3903	17.6
School after high school	31 441	36.4	8780	39.5
College degree or higher	36 865	42.7	8602	38.7
Household income					<.001
<$20 000	12 295	15.3	3081	14.7
$20 000-$34 999	18 682	23.3	5163	24.6
$35 000-$49 999	16 279	20.3	4448	21.2
$50 000-$74 999	16 343	20.4	4403	21.0
$75 000+	16 616	20.7	3887	18.5
US region					<.001
Notheast	19 912	23.1	5193	23.4
South	22 091	25.6	5387	24.2
Midwest	19 145	22.2	4700	21.1
West	25 143	29.1	6951	31.3
Marital status					<.001
Single, never married	4049	4.7	886	4.0
Married/living as married	53 919	62.6	14 476	65.3
Widowed, separated or divorced	28 117	32.7	6819	30.7
Health insurance					<.001
None	2239	2.6	810	3.7
Public insurance	9322	10.9	2003	9.1
Private insurance	45 996	53.7	13 238	60.0
Public and private insurance	26 740	31.2	5672	25.7
Other insurance not listed above	1277	1.5	349	1.6
Current smoker					0.775
No	80 987	93.9	20 876	93.9
Yes	5304	6.1	1355	6.1
Alcohol					<.001
Nondrinker	25 219	29.3	7937	35.8
<1 drink/week	27 191	31.6	6356	28.7
1 to <7 drinks/week	22 477	26.1	5859	26.4
≥7 drinks/week	11 101	12.9	2027	9.1
Any dietary supplement use	47 703	55.3	10 898	49.0	<.001
Postmenopausal hormone use					<.001
Never	34 564	40.1	9042	40.7
Past	12 841	14.9	4443	20.0
Estrogen alone	21 577	25.0	5118	23.0
Estrogen + progestin	17 231	20.0	3618	16.3
Leisure time physical activity, MET-h/week					<.001
None	11 534	13.4	3946	17.7
>0 to ≤9.5	30 411	35.2	9069	40.8
>9.5 to ≤20.5	23 565	27.3	5386	24.2
>20.5	20 781	24.1	3830	17.2

Abbreviations: CT, WHI Clinical Trials; DM-C, WHI Dietary Modification Trial Comparison Group; OS, WHI Observational Study; WHI, Women’s Health Initiative.

^a Only includes WHI participants in the analytic sample for this report.

Stage 2 calibration equations calculated metabolite dietary pattern signatures using the metabolites from NPAAS-OS and regressed these signatures on self-report dietary pattern score for each NPAAS-OS self-report instrument (FFQ, 4DFR, 24 HR) for HEI-2010 and AHEI-2010. Table 3 examined the performance of the models where we used adjusted R² (as percentage of variation explained) for HEI-2010 and AHEI-2010 using metabolomic and other nutritional biomarkers with and without participant characteristics. Adding participant characteristics showed meaningful improvement of the model performance, to adjusted R² of 57.5% (FFQ), 61.6% (4DFR), and 52.5% (24-hour recall), respectively. Similarly, adding participant characteristics significantly improved the model performance for calibration for AHEI-2010, with adjusted R² from 27.7% to 48.8% (FFQ), 47.7% to 62.6% (4DFR), and 31.0% to 53.2% (24-hour recall), respectively. To evaluate whether interaction of BMI and self-reported dietary patterns (FFQ, 4DFR, and 24-HR) should be included in the calibration equation, we plotted the analysis of covariance between calibrated dietary pattern and self-reported dietary pattern with and without BMI interaction (Figure S1). The linear patterns between self-reported HEI-2010 and the calibrated HEI-2010 demonstrated different slopes by BMI category. R² improved in the calibration equation with the interaction term, with P < .001 for interaction. Thus, the interaction between BMI and self-reported dietary scores were included in the calibration equations.

Table 3.

Calibration equation β coefficients, standard errors, and percentage of variation explained as R² from linear regression of log (Metabolomic score) on 3 dietary self-report measures (self-reported HEI-2010 score and AHEI-2010 score), Women’s health initiative nutrition and physical activity assessment study observational study (n = 450), 2006–2009.

	HEI-2010 score computed through self-report dietary measure
	FFQ (n = 406)					4DFR (n = 406)					24-HR (n = 403)
Metabolomic signature through urine and blood biomarkers	β	SE (β)	P	unadj R 2  (%)	adj b  R2  (%)	β	SE (β)	P	unadj R 2  (%)	adj b  R2  (%)	β	SE (β)	P	unadj R 2  (%)	adj b  R2  (%)
Model without covariates
Intercept	70.12	0.42				70.12	0.40				70.19	0.42
HEI-2010 self-report score	0.37	0.04	<.0001	16.8	27.1	0.35	0.03	<.0001	24.3	39.2	0.31	0.03	<.0001	18.3	29.5
Model with covariatesa
Intercept	70.59	1.27				70.33	1.24				70.36	1.36		.
HEI-2010 self-report score	0.54	0.07	<.0001	8.75	14.11	0.45	0.05	<.0001	11.71	18.89	0.34	0.05	<.0001	9.00	14.52
BMI at visit 1			<.0001	5.42	8.74			<.0001	5.01	8.08			<.0001	4.98	8.03
Normal	Ref					Ref					Ref
Overweight	−3.32	0.99				−2.83	0.98				−3.47	1.01
Obese	−4.87	0.97				−3.59	0.96				−4.82	0.98
HEI-2010 score × BMI			<.0001	11.42	18.42			0.0004	13.69	22.08			0.03	9.93	16.02
Normal	Ref					Ref					Ref
Overweight	−0.46	0.09				−0.28	0.07				−0.18	0.08
Obese	−0.26	0.09				−0.18	0.07				−0.11	0.07
Race/ethnicity			<.0001	3.67	5.92			<.0001	3.39	5.47			0.0002	3.30	5.32
White	Ref					Ref					Ref
Black	1.78	1.06				1.88	1.04				1.70	1.09
Hispanic	−3.62	1.17				−3.76	1.14				−2.94	1.20
Other race	5.65	2.21				4.79	2.16				4.60	2.26
Education			0.004	4.74	7.65			0.004	4.38	7.06			0.004	4.68	7.55
≤HS/GED	Ref					Ref					Ref
Post HS	1.26	1.21				0.56	1.18				0.26	1.24
College+	3.38	1.20				3.08	1.18				3.03	1.24
Total recreational physical activity (MET-hrs/week)	0.05	0.03	0.09	1.65	2.66
Dietary supplement use											1.60	0.78	0.04	0.66	1.06
Total R2				35.64	57.5				38.18	61.6				32.53	52.5
	AHEI-2010 score computed through self-report dietary measure
	FFQ (n = 410)					4DFR (n = 410)					24-HR (n = 407)
	β	SE (β)	P value	unadj R2 (%)	adjc  R2 (%)	β	SE (β)	P value	unadj R2 (%)	adjc  R2 (%)	β	SE (β)	P value	unadj R2 (%)	adjc  R2 (%)
Model without covariates
Intercept	53.15	0.47				53.10	0.44				53.16	0.46
AHEI-2010 score	0.34	0.04	<.0001	12.8	27.7	0.36	0.03	<.0001	21.9	47.7	0.32	0.04	<.0001	14.3	31.0
Model with covariatesa
Intercept	56.04	0.96				55.50	0.94				56.16	0.96
AHEI-2010 score	0.26	0.07	<.0001	6.25	13.59	0.25	0.056	<.0001	9.79	21.28	0.13	0.06	<.0001	6.96	15.13
BMI at visit 1			<.0001	5.13	11.15			<.0001	4.67	10.15			<.0001	5.02	10.91
Normal	Ref					Ref					Ref
Overweight	−2.63	1.17				−2.71	1.13				−3.26	1.15
Obese	−5.20	1.13				−4.26	1.10				−4.98	1.12
AHEI-2010 score × BMI			0.81	6.32	13.74			0.27	10.00	21.74			0.01	7.82	17.00
Normal	Ref					Ref					Ref
Overweight	−0.03	0.12				0.05	0.09				0.14	0.09
Obese	0.02	0.10				0.10	0.08				0.24	0.09
Race/ethnicity			0.01	3.93	8.54			0.06	3.59	7.80			0.04	3.96	8.61
White	Ref					Ref					Ref
Black	−3.16	1.25				−2.21	1.21				−2.90	1.23
Hispanic	−3.43	1.31				−2.95	1.26				−2.75	1.31
Other race	0.70	2.58				0.55	2.48				0.42	2.54
Dietary supplement use	1.69	0.90	0.06	0.80	1.74	1.83	0.86	0.03	0.73	1.59	1.86	0.89	0.04	0.72	1.57
Total R2				22.4	48.8				28.8	62.6				24.5	53.2

Abbreviations: AHEI, Alternative Healthy Eating Index; BMI. body mass index; FFQ, food frequency questionnaire; 4DFR, 4-day food record; HEI, Healthy Eating Index; NPAAS, Nutrition and Physical Activity Assessment Study; 24-HR, 24-hour recall.

Covariate selection using stepwise approach, with P value for entry and removal = 0.10. Age, total recreational physical activity, dietary score biomarker, self-reported dietary score were centered at the NPAAS sample mean.

^a Potential covariates for inclusion: race/ethnicity (White—ref; Black, Hispanic, other), age, BMI(Normal = ref, obese, overweight), interaction of BMI with Self-report Dietary Score, Education (≤HS/GED—ref, post HS, college+), season of FFQ completion (summer, spring, fall, winter = ref), current smoking, total recreational physical activity (MET-hrs/week), any dietary supplement use.

^b Adjusted for correlation between the dietary score from the NPAAS-FS and NPAAS overlap sample by dividing by 80% upper CI of the correlation coefficient (=0.67).

^c Adjusted for correlation between the dietary score from the NPAAS-FS and NPAAS overlap sample by dividing by 80% upper CI of the correlation coefficient (=0.58).

When examining the diet–cancer association, without metabolomic calibration, self-reported HEI-2010 score (via FFQ) was associated with lower risk of colorectal (Table 4; HR, 0.94; 95% CI, 0.90-0.99), lung (HR, 0.90; 95% CI, 0.86-0.94), bladder (HR, 0.86; 95% CI, 0.75-0.99), and total invasive cancer (HR, 0.98; 95% CI, 0.96-0.99) for each 20% increase (improvement) in HEI-2010 score. However, using the metabolomic calibration of the dietary self-report, some point estimates moved further away from the null value with stronger magnitudes of association (eg, colorectal and bladder cancers) but the 95% CIs included the null value of 1.0. Other findings included that a 20% increment in metabolomic-calibrated HEI-2010 was associated with 21% lower risk of lung cancer (95% CI, 0.71-0.88) and 4% lower risk of total invasive cancer (95% CI, 0.92-1.00). For models assessing associations of AHEI-2010 with cancer risk, most models were null regardless of whether the self-report was calibrated. However, one association that emerged as a higher risk of ovarian cancer with (HR, 1.48; 95% CI, 1.16-1.88) and without metabolomic calibration (HR, 1.14; 95% CI, 1.05-1.23) for each 20% improvement in AHEI-2010 score. Sensitivity analysis using IPW of inclusion to assess the diet–cancer association showed similar results for both HEI-2010 and AHEI 2010 (Table S3).

Table 4.

Cancer incidence HRs and 95% CIs for 20% increment in the dietary pattern scores, by cancer type, with and without metabolomics calibration among WHI cohorts (n = 108 522) of postmenopausal US women enrolled during 1993-1998 at 40 US clinical centers and followed through February 2023.

Cancer site	n	Without metabolomic calibration			With metabolomic calibration
		HR	95% CI	P	HR	95% CI	P
HEI-2010
Breasta	6877	1.01	0.99-1.04	.37	1.04	0.98-1.11	.21
Colorectalb	1742	0.94	0.90-0.99	.03	0.90	0.79-1.02	.10
Endometrialc	1190	1.02	0.96-1.10	.51	1.04	0.88-1.22	.68
Ovary	599	1.07	0.97-1.18	.17	1.15	0.93-1.43	.21
Leukemia	596	0.94	0.86-1.03	.18	0.87	0.72-1.07	.19
Lung	1960	0.90	0.86-0.94	<.0001	0.79	0.71-0.88	<.0001
Lymphoma	1104	1.0	0.93-1.08	.94	1.04	0.87-1.23	.70
Bladder	246	0.86	0.75-0.99	.03	0.75	0.54-1.02	.07
Kidney	419	0.91	0.82-1.00	.06	0.85	0.67-1.08	.19
Pancreas	571	1.01	0.92-1.11	.81	1.01	0.81-1.26	.93
Obesity-relateda,b,d	9809	1.0	0.97-1.02	.81	1.01	0.96-1.07	.64
Total invasivea,b	17 241	0.98	0.96-0.99	.01	0.96	0.92-1.00	.03
AHEI-2010
Breasta	6877	0.98	0.96-1.00	.11	0.95	0.88-1.02	.12
Colorectalb	1742	0.96	0.91-1.00	.07	0.9	0.78-1.05	.18
Endometrialc	1190	1.0	0.94-1.05	.89	0.97	0.82-1.14	.69
Ovary	599	1.14	1.05-1.23	.002	1.48	1.16-1.88	.002
Leukemia	596	0.97	0.90-1.04	.40	0.87	0.69-1.11	.26
Lung	1960	1.0	0.95-1.04	.90	1.01	0.88-1.16	.83
Lymphoma	1104	1.0	0.94-1.06	.90	0.98	0.82-1.17	.83
Bladder	246	1.05	0.92-1.18	.48	1.1	0.75-1.61	.63
Kidney	419	0.97	0.88-1.06	.51	0.91	0.69-1.20	.50
Pancreas	571	1.01	0.94-1.10	.73	1.02	0.79-1.30	.91
Obesity-relateda,b,d	9809	0.98	0.96-1.00	.07	0.95	0.90-1.01	.10
Total invasivea,b	17 241	0.99	0.97-1.00	.15	0.97	0.93-1.02	.22

Abbreviations: AHEI, Alternative Healthy Eating Index; BMI, body mass index; DM-C, dietary modification trial control arm; FFQ, food frequency questionnaire; HEI, Healthy Eating Index; OS, observational study; WHI, Women's Health Initiative.

HR estimates and 95% CIs are based on Cox models with baseline hazard rates stratified on study component (DM-C or OS), hormone therapy trial status (estrogen plus progestin, estrogen plus progestin placebo, estrogen alone, estrogen-alone placebo, not randomized), age at enrollment (50-54, 55-59, 60-64, 65-69, 70-74, ≥75 years), and BMI categories (<25, 25-30, ≥30), and with adjustment for a disease-specific set of potential confounding factors (including age (continuous), BMI (continuous), race/ethnicity, education, season of FFQ completion, current smoking status, any supplement use, recreational physical activity in Mets/week.

^a Adjusted for Gail Risk Score.

^b Adjusted for family history of colorectal cancer.

^c Women who had hysterectomy at baseline (WHI enrollment for OS participants and year 1 for DM-C participants) were excluded.

^d Obesity-related cancer defined here as breast, colon, rectum, endometrial, or kidney cancer.

We next evaluated the interaction between BMI and calibrated HEI-2010. We found that BMI modified the association between calibrated HEI-2010 and the risk of breast (Table 5, P = .09 for interaction), colorectal (P = .006 for interaction), leukemia (P = 0.009 for interaction), lung (P = .06 for interaction), bladder (P = .008 for interaction), and pancreatic cancers (P = .009 for interaction). For every 20% increment in calibrated HEI-2010, the colorectal cancer risk decreased by 15% (95% CI, 0.74-0.98) and 42% (95% CI, 0.39-0.85) among postmenopausal women with normal weight/underweight and overweight, respectively. But higher calibrated HEI-2010 score was not associated with risk of colorectal cancer among postmenopausal women with obesity (HR, 1.08; 95% CI, 0.87-1.33). Similar findings were observed in lung cancer, showing that improved dietary pattern (calibrated HEI-2010) was associated with lower risk of lung cancer among women with normal weight/underweight and overweight, but not among women with obesity. While, with 20% increase in calibrated HEI-2010, the risk of leukemia or bladder cancer decreased by 47% (95% CI, 0.36-0.77) and 65% (95% CI, 0.20-0.63) among postmenopausal women with obesity, but not among women with normal weight/underweight or overweight. Interestingly, higher HEI-2010 was associated with higher risk of pancreatic cancer among women with overweight (OR, 2.06; 95% CI, 1.02-4.13). We also observed that BMI modified the association between calibrated AHEI-2010 and risk of both colorectal and endometrial cancer. Specifically, with 20% increase in calibrated AHEI-2010, the risk of colorectal cancer decreased by 25% (95% CI, 0.57-0.98) only among postmenopausal women with overweight, not among women with normal weight/underweight or obesity. For endometrial cancer, 20% increase in calibrated AHEI-2010 is associated with increased risk among women with overweight (HR, 1.55; 95% CI, 1.08, 2.21), but not among women with normal weight/underweight or obesity. Sensitivity analyses showed similar results when excluding women with a BMI < 18.5 kg/m² (n = 1124, data not shown).

Table 5.

Total cancer incidence, HRs and 95% CIs for 20% increment in the calibrated dietary scores, by BMI (kg/m²) groups, with metabolomic calibration among WHI cohorts (n = 94 630) of postmenopausal US women enrolled during 1993-1998 at 40 US clinical centers and followed through February 2023.

			Metabolomic-calibrated HEI-2010		Metabolomic-calibrated AHEI-2010
		n	HR (95% CI)	P	HR (95% CI)	P
Breastc				.09a		.64b
	BMI < 25	2567	1.07 (1.00-1.16)	.06	0.94 (0.83-1.05)	.28
	BMI 25-30	2350	1.11 (0.90-1.37)	.31	0.99 (0.87-1.13)	.89
	BMI ≥ 30	1960	0.94 (0.84-1.05)	.30	0.92 (0.83-1.02)	.10
Colorectald				.006a		.07c
	BMI < 25	591	0.85 (0.74-0.98)	.03	0.81 (0.63-1.05)	.11
	BMI 25-30	602	0.58 (0.39-0.85)	.006	0.75 (0.57-0.98)	.03
	BMI ≥ 30	549	1.08 (0.87-1.33)	.49	1.05 (0.87-1.28)	.62
Endometriale				.58a		.01b
	BMI < 25	403	0.98 (0.81-1.19)	.85	0.85 (0.63-1.15)	.30
	BMI 25-30	345	0.90 (0.53-1.53)	.69	1.55 (1.08-2.21)	.02
	BMI ≥ 30	442	1.13 (0.86-1.47)	.38	0.86 (0.69-1.06)	.16
Ovary				.52a		.97b
	BMI < 25	261	1.09 (0.87-1.37)	.46	1.54 (1.06-2.24)	.02
	BMI 25-30	195	1.06 (0.56-2.02)	.86	1.47 (0.93-2.31)	.10
	BMI ≥ 30	143	1.46 (0.89-2.37)	.13	1.44 (1.00-2.08)	.05
Leukemia				.009a		.25b
	BMI < 25	249	0.99 (0.79-1.23)	.89	0.92 (0.63-1.33)	.64
	BMI 25-30	208	1.12 (0.57-2.22)	.74	1.11 (0.73-1.69)	.63
	BMI ≥ 30	139	0.53 (0.36-0.77)	.001	0.71 (0.49-1.02)	.07
Lung				.06a		.40b
	BMI < 25	844	0.78 (0.70-0.88)	<.0001	0.92 (0.75-1.13)	.41
	BMI 25-30	660	0.52 (0.35-0.75)	.0006	1.01 (0.79-1.29)	.94
	BMI ≥ 30	456	0.83 (0.66-1.05)	.13	1.12 (0.90-1.39)	.31
Lymphoma				.18a		.43b
	BMI < 25	422	1.09 (0.89-1.32)	.42	1.10 (0.82-1.48)	.52
	BMI 25-30	403	1.29 (0.79-2.10)	.32	1.05 (0.77-1.44)	.76
	BMI ≥ 30	279	0.83 (0.60-1.13)	.24	0.86 (0.65-1.14)	.28
Bladder				.008a		.20b
	BMI < 25	85	1.02 (0.68-1.53)	.92	1.66 (0.84-3.27)	.14
	BMI 25-30	99	0.49 (0.19-1.29)	.15	1.25 (0.62-2.53)	.53
	BMI ≥ 30	62	0.35 (0.20-0.63)	.001	0.78 (0.46-1.34)	.37
Kidney				.73a		.73b
	BMI < 25	108	0.88 (0.67-1.15)	.35	1.03 (0.59-1.81)	.91
	BMI 25-30	148	0.68 (0.29-1.56)	.36	0.76 (0.45-1.30)	.31
	BMI ≥ 30	163	0.76 (0.51-1.13)	.17	0.89 (0.63-1.27)	.52
Pancreas				.009a		.19b
	BMI < 25	203	1.11 (0.85-1.45)	.43	1.04 (0.69-1.57)	.84
	BMI 25-30	214	2.06 (1.02-4.13)	.04	1.35 (0.87-2.11)	.18
	BMI ≥ 30	154	0.73 (0.51-1.05)	.09	0.81 (0.57-1.17)	.26
Obesity-relatedc,d,f				.49a		.91b
	BMI < 25	3535	1.02 (0.96-1.09)	.46	0.93 (0.84-1.03)	.18
	BMI 25-30	3303	0.93 (0.78-1.11)	.44	0.96 (0.86-1.08)	.52
	BMI ≥ 30	2971	0.98 (0.89-1.07)	.64	0.94 (0.87-1.02)	.15
Total invasivec,d				.61a		.46b
	BMI < 25	6535	0.96 (0.92-1.01)	.10	0.94 (0.88-1.02)	.12
	BMI 25-30	5882	0.92 (0.81-1.05)	.21	1.01 (0.93-1.10)	.82
	BMI ≥ 30	4824	0.93 (0.86-1.00)	.05	0.96 (0.90-1.03)	.24

Abbreviations: AHEI, Alternative Healthy Eating Index; BMI, body mass index; DM-C, dietary modification trial control arm; FFQ, food frequency questionnaire; HEI, Healthy Eating Index; OS, observational study; WHI, Women's Health Initiative.

HR estimates and 95% CIs are based on Cox models with baseline hazard rates stratified on study component (DM-C or OS), hormone therapy trial status (estrogen plus progestin, estrogen plus progestin placebo, estrogen alone, estrogen-alone placebo, not randomized), and age at enrollment (50-54, 55-59, 60-64, 65-69, 70-74, ≥75 years), and with adjustment for a disease-specific set of potential confounding factors (including age (continuous), BMI categories (<25, 25-30, ≥30), interaction of dietary scores with BMI categories, Race/Ethnicity, Education, Season of FFQ completion, Current Smoking Status, Any Supplement Use, Recreational physical activity in Mets/week.

^a  P value for interaction between calibrated HEI-2010 and BMI category.

^b  P value for interaction between calibrated AHEI-2010 and BMI category.

^c Adjusted for Gail risk score.

^d Adjusted for family history of colorectal cancer.

^e Women who had hysterectomy at baseline (WHI enrollment for OS participants and year 1 for DM-C participants) were excluded.

^f Obesity-related cancer defined here as breast, colon, rectum, endometrium, or kidney cancer.

Discussion

This study used aqueous and lipid metabolomics derived from fasting blood and 24-hour urine to develop calibration equations for HEI-2010 and AHEI-2010 dietary patterns and subsequently applied the calibrated self-reported diet to disease risk models in WHI. Using this process, both HEI-2010 and AHEI-2010 satisfied the prespecified criterion of model R² ≥ 0.36, thus, calibration equations using FFQs, 4DFR, and three 24HRs from NPAAS-OS were developed. Both HEI-2010 and AHEI-2010 were found to have a reliable biomarker signature that reflect these dietary patterns, and therefore, the metabolomic-calibrated dietary pattern scores can be used for diet-disease association analyses.

In the current study, we identified 15 metabolomic and nutritional biomarkers linked to HEI-2010 and 15 linked to AHEI-2010, with 4 of those metabolites common to both dietary patterns. The common metabolites we identified, including a microbial metabolite, 2 fatty acids, and a lipid biomarker. Indole-3-propionate reflects the intake of tryptophan-rich foods, such as eggs, dairy, turkey, and legumes, and consumption of fiber, polyphenols, and fermented foods. Trans, cis-linoelaidic acid, cholesterol ester (CE 18:0), and triacylglycerol (TAG 50:3, FA 16:0) reflect dietary and endogenous saturated fat intake, such as consumption of red meat, dairy fat, butter, coconut and palm oil. Consistent with our findings, previous studies showed that certain metabolites, such as lipids, amino acids, and fatty acids, are commonly associated with different dietary patterns.^31,32 These findings suggest potential overlap in the physiological pathways influenced by different dietary patterns. Research is needed to understand these overlapping metabolomic mechanisms and their role between diet and health outcomes across various populations, with the goal to develop dietary interventions to reduce disease risk.

When the metabolomic-calibrated dietary patterns were applied to the WHI cohorts of postmenopausal women (n = 108 522) with more than 17 000 cancer cases accrued during follow-up, compared to analyses without calibration, metabolomic-calibrated HEI-2010 modestly strengthened the associations for some cancers. Specifically, metabolomic calibration of the FFQ-derived HEI-2010 scores yielded 21% and 4% lower risk for lung cancer and total invasive cancer with each 20% increment in HEI-2010 score. In contrast, without metabolomic calibration, each 20% increment in HEI-2010 yielded 10% and 2% lower risk for lung and total invasive cancers. The calibration approach may mitigate measurement errors using self-reported diet in cancer risk estimates. However, models using calibrated AHEI-2010 were mostly null except for an unexpected, apparent higher risk of ovarian cancer associated with better diet quality. Underlying mechanisms associated between higher AHEI-2010 score and higher risk of ovarian cancer are unclear. No other reports have identified dietary factors related to prevention of ovarian cancer.³³ Since ovarian cancer is usually diagnosed at an advanced stage,³⁴ it is possible that preclinical symptoms (eg, bloating, constipation), which are often misdiagnosed could initiate changes in diet and spurious results. A similar unexpected finding was observed between higher HEI-2010 score and higher pancreatic cancer risk among women with a BMI of 25 to 30 kg/m². Pancreatic cancer is also often diagnosed at a late stage.³⁵ The non-specific early symptoms, such as mild abdominal discomfort, weight loss, and fatigue, may lead to changes in dietary intake, and the confounded results. However, due to the relatively small sample size for ovarian and pancreatic cancers, these findings warrants cautious interpretation. The possibility of chance findings cannot be excluded given the multiple comparisons performed. Similar caution is urged for interpretation of obesity-stratified results for both bladder cancer and leukemia where higher BMI (≥30 kg/m²) had an inverse association with biomarker calibrated dietary patterns. The numbers of cases were much smaller in those cell sizes and it could be the result of chance.

The development of metabolomic-based dietary pattern score accounting for the difference in BMI and the estimated cancer risk associated with the physiological variations by BMI is a novel finding. Although similar dietary patterns may yield metabolomic profiles, variations may arise due to genetics, phenotypic, and environmental factors. Differences in BMI often involve different levels of dietary absorption and metabolism, thereby influencing health outcomes and disease risks.²¹ Previous studies showed that increased BMI was associated with greater absorption of glucosinolate metabolites, suggesting that BMI has a major impact on the bioavailability of dietary compounds.³⁶ Research in lipidomics and metabolomics revealed a strong link between BMI and dysregulation in lipid metabolism,^37-40 highlighting the potential impacts of obesity on various lipid metabolic pathways.^37,41 Furthermore, metabolism, influenced by diet, exercise and other exposures, can account for up to 49% of BMI variability.^38,39 Other studies demonstrated that many BMI associated metabolites were also linked with the risk of diabetes and cardiovascular disease, independent of BMI.^40-42 It is possible that the obesity-associated alterations in systemic metabolism, and the interaction of these alterations with inflammation, cell energetics, are key drivers in the development and progression of obesity-related disease, including diabetes, cardiovascular disease, and cancer.^43,44 However, further research is crucial to understand the causal pathways linking BMI-related modifications to metabolomic dietary patterns, as these mechanisms remains unclear. Investigating these pathways could inform targeted interventions, facilitating dietary recommendations that are tailored to BMI categories with implications for disease prevention.

In our study, metabolomic-calibrated dietary patterns (via FFQ) and cancer risk for some cancers (eg, colorectal and lung) differed by BMI groups, showed higher dietary pattern scores were associated with lower cancer risk for women who were normal weight or overweight, but not for those with obesity. Conversely, for other types of cancers (eg, leukemia, bladder), higher dietary pattern scores were associated with lower cancer risk only for women with obesity. The underlying mechanisms of these variations remain unclear. Nevertheless, among individuals where no association observed between higher dietary pattern score and cancer risk, other lifestyle modifications, such as increasing physical activity, improving sleep quality, limiting alcohol intake, and managing stress, may offer health benefits and lower cancer risk. Future studies are necessary to investigate how these lifestyle modifications interact with dietary patterns to impact cancer risk across various populations. These investigations could provide critical insights into targeted approaches using lifestyle interventions for cancer prevention.

Our study had several strengths. We used a controlled feeding study for the discovery phase where permitted participants maintained their usual diet, which reduced disturbances in biomarkers and revealed the metabolomic signatures of usual diet. This method allowed us to capture a more accurate representation of dietary patterns with usual diet, and the associated metabolomics profile. The use of a stringent criterion of cross-validated model R² > 0.36 essentially equaled the variance explained for gold standard biomarkers.¹⁷ While this requirement may seem arbitrary, employing benchmark biomarkers, such as DLW and urinary nitrogen, can enhance the rigor of the study. Additionally, the richness of metabolomic and nutritional data, the ability of using habitual diet feeding study to facilitate nutritional biomarker discovery, the large sample size, and the prolonged follow-up period for the WHI cohort are the major strengths of the study.⁴⁵ Lastly, cancer outcomes in the WHI were from central adjudication, and generally, there were minimal missing data of loss to follow-up.

There were some limitations of the study. The breadth of metabolomic profiles captured was restricted to the platforms used. The metabolomic signatures identified were limited to the dietary elements that comprised both HEI-2010 and AHEI-2010. It is possible that some aspects of the dietary patterns might not have been captured with the biomarker panels we tested. We recognize that some metabolites may be produced endogenously, some may come from diet, and still others maybe the downstream product of metabolism of ingested foods. We leveraged the feeding study design to develop metabolomic-based scores reflecting the foods consumed in dietary patterns. Nonetheless, we are not able to rule out that some metabolites might be the result of endogenous production. Further research into nutritional biomarker discovery is essential for the field of nutritional epidemiology. The NPAAS-OS involved nine clinical centers from the 40 WHI clinical centers, with overrepresentation of race/ethnicity minority and relative high BMI. While the NPAAS-FS participants were more homogeneous regarding demographic characteristics, and all resided in the Seattle metropolitan area. Although we have attempted to mitigate differences that may influence the generalizability of metabolomic-based equations for dietary pattern scores, some race groups are not able to be adjusted due to small numbers and there might be potential effect modification that cannot be taken into account given the limited sample size in NPAAS-OS and NPAAS-FS. It is crucial to develop nutritional biomarker-calibrated dietary patterns in other populations to have a broader implication of the results. As in any observational study, our findings are associative and may be influenced by residual confounding. There may be uncontrolled confounding that could influence the identified metabolomic and nutritional biomarkers, the calibration equations, and consequently, the estimation of cancer risk. Slightly more than 5% of participants were excluded from this analysis since their self-reported energy on the FFQ fell outside reliable ranges, thus lowering the disease association sample size slightly. Although there are some differences in baseline demographics and lifestyle characteristics between included and excluded participants (Table S4), sensitivity analyses using IPW of inclusion showed similar results (Table S3). Nevertheless, this potential selection bias may influence the generalizability of our findings.

In conclusion, we found that a panel of metabolomic and nutritional biomarkers met the prespecified criterion of model R² > 36% for HEI-2010 and AHEI-2010. In the observational study setting, the FFQ, 4DFR, and 24-hour recall also satisfied calibration equation criteria for HEI-2010 and AHEI-2010. When applied to cancer risk associations, the calibration approach mitigated measurement errors from self-reported diet (from FFQ) and provided stronger associations for some cancers. Our findings highlight the potential to use metabolomic measures for objectively measured dietary patterns and enhance studies of diet–cancer associations.

Supplementary material

Supplementary material is available at the American Journal of Epidemiology online.

Funding

The study was funded by National Cancer Institute grant K00 CA253745 (X.Z), National Cancer Institute grant R01 CA119171 (M.L.N. and J.W.L.), National Cancer Institute grant P30 CA015704, and the Women’s Health Initiative (WHI) Program, which is supported by the National Heart, Lung, and Blood Institute through contracts HHSN268201600018C, HHSN268201600001C, HHSN268201600002C, HHSN268201600003C, and HHSN268201600004C.

Conflict of interest

The authors declare no conflicts of interest.

Data availability

Procedures for obtaining permissions and access to data are available at www.whi.org.