On the use of the healthy lifestyle index to investigate specific disease outcomes

Study population

EPIC is an ongoing multicentric prospective study originally designed to study the relationship between diet and cancer risk³⁰. EPIC recruited over 500,000 men and women between 1992 and 2000 from 23 centers in 10 European countries. In our analyses we excluded participants from centres lacking information on occurrence or date of diagnosis of T2D or CVD (France, Norway, Greece and Malmö; n = 168,382), participants with no follow-up for mortality (n = 1746) or no information on lifestyle (n = 934), participants with missing information on the incidence of T2D, CVD, and/or cancer during follow-up (n = 63,842), participants with prevalent T2D, CVD, or cancer at recruitment (n = 23,864), and, for sake of simplicity, participants with missing information on any of the five variables used in the definition of HLI (n = 5786), defined as complete-case analysis.

Health-related outcomes

Data on vital status and incidence of T2D and CVD (coded using the 10th Edition of the International Classification of Diseases, ICD-10), and cancer (coded according to the International Classification of Diseases for Oncology, ICD-O-3) were collected by each participating centre, from inclusion in the study to a center- and outcome-specific last date of ascertainment^30,31,32. Dates of death were collected using record linkage with cancer registries, boards of health and death indices, or through active follow-up. Incident T2D cases, defined as E11 (ICD-10), were ascertained by a combination of self-report, linkage to primary care registers, secondary care registers, medication use (drug registers), hospital admissions, and mortality data³¹. CVD endpoints, defined as a composite of ischemic heart diseases (I20-I25), atrial fibrillation (I48), and cerebrovascular disease (I60-I69), were ascertained by different methods depending on the follow-up procedures by centre, using active follow-up through questionnaires or linkage with morbidity and hospital registries, or both³². Incident first primary cancer cases (excluding non-melanoma skin cancers) were identified through a combination of center-specific methods, including health insurance records, cancer and pathology registries and active follow-up through study participants and their next-of-kin. Follow-up for each participant and event of interest began upon inclusion in the study and ended upon the occurrence of the event, loss to follow-up, or the last date of ascertainment, whichever came first.

Assessment of lifestyle exposures at baseline

BMI (kg/m2) was derived from measured height and weight in all centers, except Oxford where it was self-reported³⁰. A validated index capturing all physical activity domains (Cambridge Index) was computed from physical activity during recreational activities and at work³³. Diet, including alcohol intake, was assessed using validated country- or center-specific dietary questionnaires designed to capture habitual consumption over the year preceding the study recruitment³⁰. To measure adherence to a healthy diet, we computed the modified relative Mediterranean Diet Score (mrMDS), a version of the original Mediterranean Diet Score incorporating vegetable oil instead of olive oil¹². To avoid redundancy with the alcohol component in the HLI, our version of mrMDS omitted alcohol intake. The remaining eight mrMDS components were measured in grams per 1000 kcal to express dietary intake as energy density¹². All dietary components were divided into country-specific tertiles and scores 0 to 2 were summed up, resulting in a final mrMDS ranging from 0 to 16 with increasing scores for healthier diets. Information on smoking status was obtained using lifestyle questionnaires³⁰, as was information on variables used for adjustments in our models, including educational attainment, menopausal status in women and the use of hormones in post-menopausal women.

Healthy lifestyle indices

Following the previous definition of the HLI used in a study of multi-morbidity in EPIC⁸, we considered HLIs that combined information on participants’ exposure to smoking, alcohol intake, diet, physical activity and adiposity. To facilitate the comparison of performance between the standard and outcome-specific HLIs, we used a binary scoring with 0/1 values reflecting unhealthy/healthy behavior for each component⁸, as displayed in Table 1. The standard HLI, ranging from 0 (unhealthiest behavior) to 5 (healthiest behavior), was defined as

To more accurately reflect the potential heterogenous relationships of each component with specific disease outcomes, outcome-specific HLIs were constructed using the same categorical scoring system. Data-driven weights were derived from the parameters of the main effects (\(w_k\)) and of the interaction terms (\(\gamma _l\)) in outcome-specific adjusted Cox models, implementing a forward selection procedure in EPIC to select relevant interaction terms among the lifestyle components. The outcome-specific HLI was defined as

with weights \(w_k^*\) and \(\gamma _l^*\) corresponding to scaled versions of \(w_k\) and \(\gamma _l\) so that outcome-specific HLIs had unit variance and larger values correspond to healthier profiles.

We also considered a more comprehensive scoring system for each variable ranging from 0 (unhealthiest) to 4 (healthiest behavior), as displayed in Table 1, again following a previous definition of the HLI used in EPIC⁸.

Cox models

In all our analyses, Cox models used age as the main time scale and were stratified by study center, sex, and age at recruitment in 5-year categories. They were adjusted for education level (no schooling, primary, secondary, and university or more), height (continuous), and energy intake from non-alcoholic sources (kcal/day), and, for women, menopausal status (pre-menopausal, peri-menopausal, post-menopausal, surgical) and use of postmenopausal hormones (never, ever, unknown). For each outcome, one Cox model was constructed with all five score variables as the main exposures to derive the outcome-specific weights and the outcome-specific HLIs. Then, Cox models were constructed by considering, in turn, each version of the HLI as the main exposure. The HLI was consistently modelled in continuous using a linear term on the log-hazard-rate scale.

Evaluation criteria

HR estimates and discriminatory power

For each event, HR estimates and corresponding 95% confidence intervals (CIs) were computed for a 1-standard deviation (SD) increase of the different versions of the HLI. They allowed the comparison of the estimated association between overall adherence to a healthy lifestyle and the event being studied, depending on the version of the HLI being used. To further illustrate how risk stratification may be hindered when using an unweighted rather than an outcome-specific weighted HLI, we considered the 2⁵ = 32 lifestyle profiles corresponding to each possible combination of the five binary scores Smoking^(0–1), Alcohol^(0–1), Diet^(0–1), PA^(0–1) and Adipo^(0–1). Setting the unhealthiest profile Smoking^(0–1) = 0, Alcohol^(0–1) = 0, Diet^(0–1) = 0, PA^(0–1) = 0, Adipo^(0–1) = 0 as the reference, we compared the HR estimates for the other 31 profiles produced by Cox models utilizing the standard and outcome-specific HLIs, respectively.

More generally, the discriminatory power of models based on the different versions of the HLI was quantified using Harrell’s C-index. HRs and Harrell’s C-indices were primarily computed in the full EPIC study population. For models based on outcome-specific HLIs, this amounted to evaluating them on the data used for their construction, which could create bias if overfitting was present. Cross-validation was applied to assess this bias: the EPIC study population was randomly split into (i) a training sample (75% of the total sample) where the outcome-specific weights were estimated, and (ii) a test sample (the remaining 25% of the total sample) where HRs and Harrell’s C-indices were computed. This process was repeated 10 times to prevent possible dependency on a single split³⁴. HRs and Harrell’s C-indices were averaged over these 10 repetitions and compared to the values obtained on the total EPIC study population to assess the bias magnitude.

Population attributable fractions

For each specific outcome, we computed PAFs at age a, defined as

Here, \(P\left(Y<a\right)\) is the event risk before age a in the EPIC study population and \(P\left(Y^\left(max\right)<a\right)\) is the hypothetical event risk before age a in the counterfactual EPIC study population where, for all participants, all five lifestyle summary variables would have been set to their maximal possible values, while all other variables used for adjustment or stratification would have been set to their actual value observed in EPIC. Under technical conditions³⁵, \(PAF\left(a\right)\) coincides with the proportion \(P\left(Y^\left(max\right)>a| Y<a\right)\) of events before age a that would have been prevented had all EPIC participants adhered to the “healthiest” behavior regarding all five lifestyle components. Absolute risks \(P\left(Y<a\right)\) and counterfactual absolute risks \(P\left(Y^\left(max\right)<a\right)\) were estimated by averaging the individual risk predictions in the EPIC study population, and in the counterfactual populations, respectively. Non-parametric bootstrap based on 100 bootstrapped samples was used to estimate the corresponding 95% CI.

All analyses were performed using the R software, version 4.1.2. Given the nature of the weights used in the definition of the outcome-specific HLIs, models utilizing individual lifestyle scores would achieve similar discriminatory power and produce similar PAF estimates when compared to models based on outcome-specific HLIs.

Ethics

The EPIC study was conducted according to the Declaration of Helsinki and approved by the ethics committee at the International Agency for Research on Cancer (IARC) on 12 January 1995 and on 10 May 2017 (re-evaluation). Written informed consent was obtained from all subjects involved in the study.