A Genetic Risk Score to Personalize Prostate Cancer Screening, Applied to Population Data

Prostate cancer is the second most common malignancy in men worldwide with nearly 1.3 million cases diagnosed globally in 2018 (1). It was the third leading cause of European male cancer–related mortality in 2018, following mortality from lung and colorectal cancers (2). Prostate cancer screening with PSA testing can reduce mortality (3), but universal screening may cause overdetection of cancers that would never become clinically apparent in a man's life-time and overtreatment of indolent disease. Guidelines recommend that individual men participate in informed decision-making about screening, taking into account factors such as their age, race/ethnicity, family history, and preferences (4–6).

Assessment of a man's genetic risk of developing prostate cancer has promise for guiding individualized screening decisions (7, 8). We previously developed and validated a polygenic hazard score (PHS), a weighted sum of 54 SNP genotypes, as significantly associated with age at diagnosis of clinically significant prostate cancer, defined as cases where any of the following is applied: Gleason score ≥7, clinical stage T3–T4, PSA ≥10, or where there were nodal or distant metastases (9). Risk stratification by the PHS also improved the screening performance of PSA testing; the positive predictive value (PPV) of PSA testing for clinically significant prostate cancer increased as PHS increased (9).

Here, we apply the prostate cancer PHS to population data to assess its potential impact on individualized screening. Specifically, we combine genetic risk, measured by PHS, and known population incidence rates to estimate a risk-equivalent age, for example, the age at which a man with a given PHS will have the same risk of clinically significant prostate cancer as a typical man at age 50 years. Such genetic risk estimates can guide individualized decisions about whether, and at what age, a man might benefit from prostate cancer screening.

Full methodologic details of the development and validation of the prostate cancer PHS have been described previously (9). Briefly, the PHS was developed using PRACTICAL consortium clinical and genetic data from 31,747 men of European ancestry as a continuous survival analysis model (10) and found to be associated with age at prostate cancer diagnosis (9). Validation testing was performed in an independent, separate dataset consisting of 6,411 men from the United Kingdom ProtecT study (11, 12). PHS was calculated as the vector product of a patient's genotype (X_i) for n selected SNPs and the corresponding parameter estimates (β_i) from a Cox proportional hazards regression (Eq. 1):

The 54 SNPs included in the model, and their parameter estimates, have been published previously (9) and are also shown in Supplementary Table S1.

Age-specific prostate cancer incidence data were obtained for men ages 40–70 years from the United Kingdom, 2013–2015 (Cancer Research United Kingdom; ref. 13). Men may be less likely to be screened outside this age range (3, 14). The log of the prostate cancer incidence data were fit using linear regression to develop a continuous model of age-specific prostate cancer incidence in the United Kingdom (I_all).

The United Kingdom age-specific proportion of incidence classified as clinically significant prostate cancer was estimated using data from the Cluster Randomized Trial of PSA Testing for Prostate Cancer (CAP). The CAP trial evaluated the impact of a single, low-intensity PSA screening intervention on prostate cancer–specific mortality in the United Kingdom (15). CAP was linked to the ProtecT study, which included men ages 50–69 at randomization (15); ProtecT compared management options including surgery, radiotherapy, and active surveillance in patients with PSA-detected prostate cancer (12). The clinical and demographic features of the CAP and ProtecT studies have been described previously (12, 15). Clinically significant prostate cancer was defined as cases often ineligible for active surveillance (consistent with the definition used in the PHS development). These are cases with Gleason score ≥7, clinical stage T3–T4, PSA ≥10, or with nodal/distant metastases (9, 16, 17). Men in the intervention arm of the CAP trial who were diagnosed with any prostate cancer were divided into 5-year age intervals at prostate cancer detection (n = 8,054; ref. 15). The proportion of clinically significant disease in each age interval was calculated as the number of clinically significant prostate cancer diagnoses divided by the total number of prostate cancer diagnoses in the CAP cohort for whom PSA and clinical stage information were available (n = 6,388; ref. 15). The total (all ages) proportion of clinically significant prostate cancer was similarly calculated from CAP data. The age-specific prostate cancer incidence curve, I_all, was multiplied, within each 5-year age range, by the corresponding age-specific proportion of CAP clinically significant prostate cancer diagnoses, to yield a continuous estimate of age-specific, clinically significant prostate cancer incidence (I_{clinically significant}). A similar calculation was done to estimate age-specific, more aggressive prostate cancer incidence [using a stricter definition of clinically significant disease that corresponds to clinical high risk or above by National Comprehensive Cancer Network (NCCN) guidelines: clinical stage T3–T4, PSA >20, Gleason score ≥8, or with nodal/distant metastases; refs. 9, 16, 17] as I_{more-aggressive}. Finally, clinically insignificant prostate cancer incidence (I_{clinically insignificant}) was estimated as the difference between I_all and I_{clinically significant}.

Men in the ProtecT study with genotype data (n = 6,411) were categorized by their PHS percentile ranges (0–2, 2–10, 10–30, 30–70, 70–90, 90–98, and 98–100) to correspond to percentiles of interest (1, 5, 20, 50, 80, 95, and 99, respectively). These percentiles refer to the distribution of PHS in the ProtecT dataset within controls aged <70. Incidence rates of clinically significant prostate cancer were calculated for each percentile range (I_percentile) using Cox proportional hazards regression (parameter estimate, β), following the methods published previously (9). The reference for each HR was taken as the mean PHS among those men with approximately 50th percentile for genetic risk (i.e., 30th–70th percentile of PHS, called PHS_median), and this median group was assumed to have an incidence of clinically significant disease matching the overall population (I_{clinically significant}, calculated above). Incidence rates for the other percentiles of interest (I_percentile) were then calculated by determining the mean PHS among men in the corresponding percentile range (called PHS_percentile) and applying Eq. 2:

As described in the original validation of this PHS model for prostate cancer (9), PHS calculated in the ProtecT dataset will be biased by the disproportionately large number of cases included, relative to incidence in the general population. Leveraging the cohort design of the ProtecT study (11), we therefore applied a correction for this bias, using previously published methods (18) and the R “survival” package (R version 3.2.2; refs. 19, 20). The corrected PHS values were used to update PHS_percentile and PHS_median used in Eq. 2. Then, 95% confidence intervals for the HRs for each percentile were determined by bootstrapping 1,000 random samples from the ProtecT dataset, while maintaining the same number of cases and controls from the original dataset. The I_percentile, predicted partial hazard (product of PHS_percentile and the estimated β), and SEs (to account for sample weights) were calculated for each bootstrap sample.

Percentile-specific incidence estimates (I_percentile) were visualized as the corresponding cumulative incidence curves for clinically significant prostate cancer diagnosis for men ages 50–70 years. Analogous HRs and incidence curves were similarly calculated for the annualized incidence rates of clinically insignificant and more aggressive prostate cancer.

An individualized PHS to aid prostate cancer screening decisions in the clinic might be facilitated by a readily interpretable translation of PHS to terms familiar to men and their physicians. The PHS was, therefore, combined with United Kingdom clinically significant prostate cancer incidence data to give a risk-equivalent age: when a man with a given PHS percentile would have the same risk of clinically significant prostate cancer as, say, that of a typical man at 50 years old (50-year-standard risk). We defined ΔAge as the difference between age 50 and the age when prostate cancer risk matches that of a typical 50-year-old man. 95% confidence intervals for the age when a man reaches 50-year-standard risk and ΔAge were determined using the HRs calculated from the 1,000 bootstrapped samples from ProtecT, described above.

Finally, we considered the common clinical scenario of a man presenting to his primary care physician to discuss prostate cancer screening. To illustrate how PHS might influence this discussion, we identified the subset of men in the ProtecT validation dataset who were around the median age of 60 years (55–64), to represent a typical patient. From this subset, we created three groups: those whose prostate cancer risk-equivalent age remained within the selected range (ages 55–64), those whose risk-equivalent age was <55, and those whose risk-equivalent age was ≥65. We then calculated the PPV and SE of the mean of PSA testing for development of clinically significant prostate cancer in these three PHS-adjusted (prostate cancer risk–equivalent age) groups using methods described previously (9). This was done by taking 1,000 random samples (with replacement) of the subjects with elevated PSA (≥3.0 ng/mL) in the dataset, stratified to ensure each random sample matched the distribution of controls and cases reported for men with elevated PSA in ProtecT (11, 12). Stratification was also used to ensure the proportion of clinically significant cases matched the proportion reported in CAP for the age range of 55–64 (11), such that the PPV for the sample exactly matched the expected value for the linked ProtecT and CAP trials, but the distribution of genetic risk (PHS) was varied at random within each disease status group (control, clinically significant, and clinically insignificant). A similar calculation for PPV of PSA testing for development of any prostate cancer was performed for the three PHS-adjusted age groups.

Linear regression yielded a model of prostate cancer age–specific incidence rates (Eq. 3; R² = 0.96; P = 0.001) that was highly consistent with empirical data reported by Cancer Research United Kingdom (Fig. 1).

In the CAP study (15), the overall proportion of prostate cancer incidence classified as clinically significant disease was 72.3%. The proportions of age-specific, clinically significant disease increased with age: 48.0%, 55.9%, 63.5%, and 79.7% of men ages 50–54, 55–59, 60–64, and 65–69, respectively, were diagnosed with clinically significant prostate cancer. Combining men ages 55–64, the proportion of age-specific, clinically significant prostate cancer was 61.1%.

Cumulative incidence estimates of clinically significant prostate cancer are shown in Fig. 2 for various levels of genetic risk, as indicated by PHS percentile, showing a difference in age at diagnosis related to PHS strata. Supplementary Figs. S1 and S2 show analogous results for the incidence curves of clinically insignificant and more aggressive prostate cancer, respectively. Table 1 shows risk-equivalent age for each PHS percentile. The expected age at clinically significant prostate cancer diagnosis differs by 19 years between the 1st and 99th PHS percentiles. Specifically, a man with a PHS in the 99th percentile reached a prostate cancer detection risk equivalent to the 50-year standard at an age of 41 years. Conversely, a man with a PHS in the 1st percentile would not reach the 50-year-standard risk level until age 60 years. Qualitatively, the curves for clinically significant (Fig. 2), clinically insignificant (Supplementary Fig. S1), and more aggressive (Supplementary Fig. S2) prostate cancer maintain consistent horizontal shifts relative to curves for other PHS percentiles over the age range studied. Quantitatively, this was confirmed by ΔAge, which remained the same for each PHS percentile across a true age range of 40–70. Thus, ΔAge was taken to be approximately constant for each PHS percentile and is reported in Table 1.

Figure 3 shows the PPV of PSA testing for clinically significant prostate cancer was 0.21 (SE, 0.01) for men approximately 60 years old (data derived from a total of 1,395 ProtecT men ages 55–64: 283 with clinically significant prostate cancer, 127 with clinically insignificant prostate cancer, and 575 controls with a PSA ≥3.0 ng/mL). PPV was lower for those with a prostate cancer risk-equivalent age <55 years (0.12; SE, 0.04) and higher for those with prostate cancer risk-equivalent age ≥65 years (0.40; SE, 0.03).

The PPVs of PSA testing for any prostate cancer were 0.18 (SE, 0.05), 0.37 (SE, 0.01), and 0.61 (SE, 0.03) in men with a prostate cancer risk-equivalent age <55 years, between 55 and 64 years, and ≥65 years, respectively. These PPVs, in combination with the PPVs of PSA for clinically significant prostate cancer, indicate that in the older prostate cancer risk-equivalent age group (≥65 years), 40% of positive PSA tests are from clinically significant disease, 21% are from clinically insignificant disease, and 39% are false positives. The false positive rates for men with a prostate cancer risk-equivalent age <55 years and between 55–64 years are 82% and 63%, respectively.

We applied the PHS to population incidence data to estimate age-specific risk of clinically significant prostate cancer. The resulting age-specific incidence rates (displayed as incidence curves in Fig. 2) demonstrate clinically meaningful differences across various levels of genetic risk, as estimated by PHS. By combining these population curves with an individual's genetic risk and true age, we demonstrate calculation of a risk-equivalent age at diagnosis of clinically significant prostate cancer. This age relates a man's current prostate cancer risk to that of the age-specific population average. The incidence curves for clinically significant prostate cancer are modulated by 19 years between the 1st and 99th percentiles of PHS. Moreover, the PPV of PSA testing in three PHS-adjusted (prostate cancer risk-equivalent age) groups demonstrated that PPV is significantly higher in men with higher risk-equivalent ages of prostate cancer diagnosis. These results have important implications for clinicians considering discussions of whether, and when, to initiate prostate cancer screening in an asymptomatic man.

Prostate cancer can cause considerable mortality and morbidity, but is curable if detected early. Determination of age of clinically significant disease diagnosis is thus highly relevant. Data from the CAP study shown here confirm prior findings of increasing risk of clinically significant prostate cancer as men age (21–24). The proportion of new prostate cancer diagnoses classified as clinically significant in CAP is higher than some older studies that were limited to men with low PSA and normal digital rectal exam (25–27), while another modern population study shows similar or higher proportions with clinically significant disease (21). Taken together, these results suggest that screening delayed to an older age will yield a higher incidence of clinically significant disease.

The primary screening tool, PSA testing, is associated with a small absolute decreased risk of death from prostate cancer (3), but carries a risk of overdetection and harm from overtreatment in men who would never have experienced clinical manifestations of their prostate cancer (28). Thus, universal screening comes at a high cost, both in burden on health care systems and in the sequelae arising from elevated PSA in men with indolent disease: unnecessary biopsy procedures, overdetection, and treatment-related morbidities (4, 5). Conversely, there are some men who will develop clinically significant prostate cancer and would benefit from screening, possibly even at a relatively young age. Screening guidelines recommend individualized decision-making, but the available quantitative or objective data to guide these decisions are insufficient. For instance, family history provides some guidance, but genetic risk has been shown to be more strongly associated with age of clinically significant prostate cancer diagnosis than patient-provided family history (9, 29).

PHS, in conjunction with other informative factors such as family history, may help identify men who may develop the highest risk cancers (12). Incorporating a risk-adjusted age in an electronic medical record could reduce burden for general practitioners. The risk-adjusted age can be based on whatever threshold of risk for clinically significant prostate cancer is considered optimal. Here, we have used the typical risk at age 50. Waiting until the man whose risk-adjusted screening age reached 60 would be much more likely to avoid overdiagnosis and overtreatment than to miss a clinically significant prostate cancer. This is supported by the clinically significant–specific incidence rates reported here for CAP in the United Kingdom and also by recently reported absolute age–specific incidence rates in Norway (21). One way a risk-stratified approach addresses overdetection is by providing a quantifiable, objective, and accurate rationale to not screen many men until they reach sufficient risk (in which time, their competing risks also have a chance to manifest; these could also inform screening and management decisions, especially if they affect life expectancy). The concern for overtreatment is also a critical consideration. As demonstrated in the ProtecT study, lower risk disease does not need to be treated aggressively at diagnosis and can be monitored with active surveillance and routine PSA checks (12). In addition, other major trials have demonstrated that the risks of biopsy can be mitigated by using multiparametric prostate MRI (30–32). These important mitigating factors are not directly related to polygenic risk, but they do decrease the risks associated with a prostate cancer screening program.

The stratification of men based on their genetic risk is of particular interest in the primary care setting, where the majority of prostate cancer screening discussions take place. Shared decision-making between patient and physician has long been recommended in discussions of prostate cancer screening (5, 33), and physicians are tasked with determining an individual's risk based on factors such as his family history and ethnicity. However, physicians demonstrate different attitudes toward screening, with some screening all men proactively to avoid underdiagnoses, some screening only those men who request it, and some who attempt to weigh the costs and benefits of PSA screening on a case-by-case basis (34, 35). General practitioners, who are already limited by time constraints and their patients' other health issues, must carefully discuss the complex risks and benefits of PSA screening with their patients (36). However, efficiently identifying men at higher risk of clinically significant disease is important because detection of prostate cancer at an early stage allows for definitive treatment to prevent cancer progression or metastases (12).

Quantitative risk stratification could guide physicians in their screening conversations with patients by providing an objective risk-equivalent age for the development of clinically significant disease. This allows for simpler and more standardized informed decision-making regarding whether an individual man might benefit from prostate cancer screening. For example, physicians who normally initiate screening discussions at some age (e.g., 50–55) could shift the timing according to the prostate cancer risk-equivalent age. Some men might need to begin prostate cancer screening at a younger age to detect early onset of clinically significant disease. The PHS has previously demonstrated high PPV of PSA testing for clinically significant prostate cancer in men with progressively higher scores (9).

The potential utility of prostate cancer risk-equivalent age in the clinic is additionally demonstrated by its impact on PPV of PSA testing for clinically significant prostate cancer. Suppose a 60-year-old man presents to his physician to inquire about prostate cancer screening. If this man has a prostate cancer risk-equivalent age close to his true age (55–64), the PPV of a PSA test (for prediction of clinically significant prostate cancer) for him will be approximately 24%. If his risk-equivalent age is <55, the PPV decreases to 13%, and he might be reassured in foregoing PSA testing. Postponing, or even forgoing, screening in men with low PHS percentiles to when they reach their risk-equivalent age could decrease the harms associated with screening, or early detection and treatment of prostate cancer (4, 5). Other men may choose to delay the initiation of PSA testing until they are older and have increased risk. Conversely, if this same man has a risk-equivalent age ≥65, the PPV of PSA testing increases substantially to 45%, implying that screening may be more informative for him. Of note, the increase in PPV in this illustration exceeds that of the reported effect of carrying a mutation in BRCA1 or BRCA2 (37).

Cost-effectiveness is another concern regarding prostate cancer screening. Use of PHS, a one-time test valid for a man's entire life, can improve screening efficiency while reducing overall costs. The genotyping chip assay requires only a saliva sample and can be run for costs similar to those for single-gene testing (e.g., the BRCA mutation). Genotyping also informs genetic risks for other diseases, possibly allowing multiple tests to be run on the same genotype results (38, 39). PSA screening (and subsequent prostate biopsy) could be offered only to those men at higher risk of clinically significant disease. PHS might increase the efficiency of any prostate cancer screening program by incorporating knowledge that there are some men with higher baseline genetic risks of developing clinically significant prostate cancer, even at a younger age, while others have a low baseline genetic risk.

Limitations of this work include that the PHS did not incorporate genotypic data from men of non-European ancestry during its development (9), a reflection of the available data, which may affect the potential use of the PHS for screening decision-making in men from other ethnic groups. This is noteworthy, as disparities in prostate cancer incidence and survival show that in the United States, men with African ancestry are more likely to develop prostate cancer and to die from their disease (40). Our group and others are studying the application of genetic scores to non-European ancestry groups. In addition, we used incidence data from a single country (the United Kingdom) with relatively low rates of screening. While the epidemiologic data used in this work are of high quality and drawn from the same United Kingdom population as was previously used for the validation of the PHS model (9), further work should evaluate the PHS in other populations. Finally, there are now over 140 SNPs reported to have associations with prostate cancer, identified using a meta-analysis that included ProtecT data (41), but not all of these SNPs are represented on the custom array used to develop the original PHS. Furthermore, the PHS model was validated using independent data from ProtecT; the inclusion of those other SNPs associated with prostate cancer would have introduced circularity into the validation. Adding more SNPs to further improve the model is an area of active investigation. If we, or others, succeed in developing a further optimized PHS, we expect the range of ΔAge to expand.

We conclude that clinically meaningful risk stratification can be achieved through application of a PHS that is associated with age at clinically significant prostate cancer diagnosis to United Kingdom population data. PHS can also be used to calculate estimates of risk-equivalent age for the development of clinically significant prostate cancer for individual men. The PPV of PSA was higher for men with higher PHS-adjusted prostate cancer–equivalent ages. Assessing personalized genetic risk via PHS could assist patients and physicians, alike, with the important decision of whether, and when, to initiate prostate cancer screening.

C.C. Fan is a data scientist at HealthLytix. F.C. Hamdy is an editor-in-chief at British Journal of Urology International and has a consulting/advisory board relationship with Intuitive. R.A. Eeles reports receiving speaker honoraria from the GU-ASCO meeting (January 2016), RMH FR meeting (November 2017; support from Janssen), and University of Chicago (May 2018) and educational honorarium paid by Bayer & Ipsen to attend GU Connect at a venue at ESMO. O.A. Andreassen is a consultant for HealthLytix. A.M. Dale reports receiving a commercial research grant from GE Healthcare, has ownership interest (including patents) in CorTechs Labs, Inc., and has a consulting/advisory board relationship with Human Longevity, Inc. T.M. Seibert is a consultant for Multimodal Imaging Services Corporation and WebMD, Inc. No potential conflicts of interest were disclosed by the other authors.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH, NIHR, or the Department of Health and Social Care.

Conception and design: M.-P. Huynh-Le, Z. Kote-Jarai, T.L.J. Tammela, A.S. Kibel, O.A. Andreassen, A.M. Dale, T.M. Seibert

Development of methodology: M.-P. Huynh-Le, C.C. Fan, R. Karunamuni, C.K. Slavov, A.M. Dale, T.M. Seibert

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M.-P. Huynh-Le, E.I. Walsh, E.L. Turner, J.A. Lane, R.M. Martin, D.E. Neal, J.L. Donovan, F.C. Hamdy, D.F. Easton, Z. Kote-Jarai, K. Muir, H. Grönberg, F. Wiklund, M. Aly, J. Schleutker, C. Sipeky, T.L.J. Tammela, B.G. Nordestgaard, T.J. Key, R.C. Travis, P.D.P. Pharoah, N. Pashayan, K.-T. Khaw, S.N. Thibodeau, S.K. McDonnell, C. Maier, W. Vogel, M. Luedeke, K. Herkommer, A.S. Kibel, C. Cybulski, W. Kluzniak, L.A. Cannon-Albright, H. Brenner, B. Schöttker, B. Holleczek, J.Y. Park, T.A. Sellers, C.K. Slavov, R.P. Kaneva, V.I. Mitev, J. Batra, J.A. Clements, A.B. Spurdle, M.R. Teixeira, P. Paulo, S. Maia, H. Pandha, A. Michael, T.M. Seibert

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M.-P. Huynh-Le, R. Karunamuni, J.K. Parsons, A. Amin Al Olama, T.L.J. Tammela, N. Pashayan, D.J. Schaid, H. Brenner, J. Batra, O.A. Andreassen, A.M. Dale, T.M. Seibert

Writing, review, and/or revision of the manuscript: M.-P. Huynh-Le, C.C. Fan, R. Karunamuni, E.L. Turner, J.A. Lane, R.M. Martin, D.E. Neal, J.K. Parsons, R.A. Eeles, D.F. Easton, Z. Kote-Jarai, A. Amin Al Olama, K. Muir, J. Schleutker, C. Sipeky, T.L.J. Tammela, B.G. Nordestgaard, T.J. Key, R.C. Travis, P.D.P. Pharoah, N. Pashayan, K.-T. Khaw, S.K. McDonnell, D.J. Schaid, A.S. Kibel, C. Cybulski, H. Brenner, B. Schöttker, B. Holleczek, J.Y. Park, T.A. Sellers, H.-Y. Lin, R.P. Kaneva, V.I. Mitev, J. Batra, J.A. Clements, A.B. Spurdle, M.R. Teixeira, H. Pandha, I.G. Mills, O.A. Andreassen, A.M. Dale, T.M. Seibert

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E.I. Walsh, J.L. Donovan, S. Benlloch Garcia, K. Muir, F. Wiklund, K.-T. Khaw, A.S. Kibel, D. Wokolorczyk, W. Kluzniak, J.Y. Park, R.P. Kaneva, V.I. Mitev, A.B. Spurdle, M.R. Teixeira

Study supervision: T.L.J. Tammela, C.K. Slavov, O.A. Andreassen, T.M. Seibert

Other (coordinator of trial that contributed data): E.L. Turner

This study was funded in part by a grant from the United States National Institute of Health/National Institute of Biomedical Imaging and Bioengineering (#K08EB026503 to T.M. Seibert), United States Department of Defense (#W81XWH-13-1-0391 to A.M. Dale and T.M. Seibert), the Research Council of Norway (#223273 to O.A. Andreassen), KG Jebsen Stiftelsen (to O.A. Andreassen), and South East Norway Health Authority (to O.A. Andreassen). The CAP trial was funded by grants C11043/A4286, C18281/A8145, C18281/A11326, C18281/A15064, and C18281/A24432 from Cancer Research UK. The UK Department of Health, National Institute of Health Research provided partial funding. R.M. Martin was supported by a Cancer Research UK Programme Grant, the Integrative Cancer Epidemiology Programme (C18281/A19169), and the National Institute for Health Research (NIHR) Bristol Biomedical Research Centre. Funding for the PRACTICAL consortium member studies is detailed in the Supplementary Data.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.