A risk prediction model for head and neck cancers incorporating lifestyle factors, HPV serology and genetic markers

Abstract

Head and neck cancer is often diagnosed late and prognosis for most head and neck cancer patients remains poor. To aid early detection, we developed a risk prediction model based on demographic and lifestyle risk factors, human papillomavirus (HPV) serological markers, and genetic markers. A total of 10,126 head and neck cancer cases and 5,254 controls from 5 North American and European studies were included. HPV serostatus was determined by antibodies for HPV16 early oncoproteins (E6, E7) and regulatory early proteins (E1, E2, E4). The data were split into a training set (70%) for model development and a hold-out testing set (30%) for model performance evaluation, including discriminative ability and calibration. The risk models including demographic, lifestyle risk factors and polygenic risk score showed a reasonable predictive accuracy for head and neck cancer overall. A risk model that also included HPV serology showed substantially improved predictive accuracy for oropharyngeal cancer (AUC=0.94, 95%CI=0.92–0.95 in men and AUC=0.92, 95%CI=0.88–0.95 in women). The 5-year absolute risk estimates showed distinct trajectories by risk factor profiles. Based on the UK Biobank cohort, the risks of developing oropharyngeal cancer among 60 years old and HPV16 seropositive in the next 5 years ranged from 5.8% to 14.9% with an average of 8.1% for men, 1.3% to 4.4% with an average of 2.2% for women. Absolute risk was generally higher among individuals with heavy smoking, heavy drinking, HPV seropositivity, and those with higher polygenic risk score. These risk models may be helpful for identifying people at high risk of developing head and neck cancer.

Introduction

Head and neck cancer comprises of tumors originating in the oral cavity, hypopharynx, oropharynx, nasopharynx and larynx. In 2020, an estimated 878,348 individuals developed head and neck cancer worldwide, including 65,630 in the United States^{1, 2}. The prognosis of head and neck cancer varies by anatomical site and stage at diagnosis. While the 5-year survival rates range from about 50% to 90% for those who are diagnosed at an early stage, patients with head and neck cancer are often diagnosed at an advanced stage, in which case only 20%–40% survive past 5 years^3–5. In addition, it is well documented that patients with head and neck cancer, particularly advanced stage disease, suffer from significant psychological impact. This is a result of visible disfigurement and disruption of essential functioning due to the disease itself or the treatment^{6, 7}. Therefore, early detection and prevention of head and neck cancer is of critical importance.

Tobacco smoking and alcohol consumption are the two well-recognized risk factors for head and neck cancer^{5, 8}. Previous pooled analyses have reported more than a two-fold increased risk of head and neck cancer in cigarette smokers and infrequent alcohol drinkers compared to non-users of these substances^{9, 10}. Besides tobacco and alcohol use, infection with high-risk HPV types is also an independent risk factor of head and neck cancer. In particular HPV16 is considered a causative agent of head and neck cancer, specifically for cancers of the oropharyngeal region^{5, 11–13}. Seropositivity for HPV16 E6 is a highly sensitive and specific marker for HPV-driven oropharyngeal cancer, and blood-based HPV16 E6 antibodies can be found several years before cancer diagnosis^14–17. In addition, recent genome wide association studies have identified several genetic loci associated with head and neck cancer risk^{18, 19}.

As a tool to facilitate risk stratification, risk prediction models have been developed using known or potential risk factors for various cancers. Although there were limited previous work reported for risk prediction of head and neck cancer^20–23, none of the existing risk prediction models considered all potential major risk factors, including HPV seropositivity and genetic susceptibility^{20, 21}. In this study, we aimed to develop a prediction model that incorporated demographic, lifestyle, HPV, and identified genetic factors, and to estimate the absolute 5-year risk of developing head and neck cancer, oral cavity cancer and oropharyngeal cancer.

Methods and Materials

Study participants

Five studies from the United States, Canada and Europe were included in this analysis, with a total of 15,380 study participants, including 10,126 head and neck cancer cases and 5254 controls (Supplementary Figure 1) from the NIH-funded VOYAGER (Human Papillomavirus, Oral and Oropharyngeal Cancer Genomic Research) program. The five participating studies are Carolina Head and Neck Cancer Epidemiology (CHANCE) and Pittsburgh in the United States, Mount Sinai Hospital-Princess Margaret (MSH-PMH) study in Toronto, Canada, Alcohol-Related Cancers and Genetic susceptibility in Europe (ARCAGE), and Head and neck 5000 (HN5000) in the United Kingdom. The details of these studies have been described previously^24–28. Briefly, four of the studies are case-control in design and HN5000 is a prospective clinical cohort study with longitudinal follow up of head and neck cancer cases. All cases were patients with squamous cell carcinoma of the head and neck confirmed by pathology reports. Controls were individuals without cancer diagnosis randomly selected from the general population^{25, 28}, or the visitors of the participating hospitals^{24, 27}, often frequency-matched to cases in terms of age and sex. All participants were administered a structured questionnaire which assessed information regarding demographic, lifestyle, and medical history. Plasma samples were obtained at the time of diagnosis and prior to start of treatment for cancer cases, and at time of enrollment for controls.

HPV serology assay and genotyping

HPV antibodies were measured in oropharyngeal cancer cases and controls using a bead-based multiplex serology assay^{16, 29}. Antigens were affinity-purified, bacterially expressed fusion proteins with N-terminal glutathione S-transferase. We measured antibodies against the early oncoproteins (E6, E7) and regulatory early proteins (E1, E2, E4) for HPV16, and the antibody values were dichotomized based on predefined median fluorescence intensity (MFI) values¹⁶. We applied two criteria to determine seropositivity: 1) high antibody levels against HPV 16 E6 alone (>1000 MFI); or 2) seropositivity against three of four HPV16 early proteins (E1: >200 MFI, E2: >679 MFI, E6: >484 MFI and E7: >548 MFI). Participants were considered HPV seropositive if either of these 2 criteria was met³⁰. HPV serology was performed at the German Cancer Research Center (DKFZ, Heidelberg, Germany), and laboratory personnel were blinded to the disease status. For controls that were not assayed, we imputed their serostatus by random binomial draw with the overall probability of seropositivity (0.86%) estimated from controls who were assayed^{31, 32}.

The genetic risk variants included in the model were those previously identified in the genome wide association studies of upper aerodigestive tract cancer risk^{18, 19, 33, 34}. A total of 22 variants were included and are summarized in the Supplementary Table 1, including 10 for head and neck cancer overall, 5 for oral cavity, and 10 for oropharyngeal cancer. The genotype data for variants were extracted from the head and neck cancer OncoArray dataset previously published¹⁸. We computed a polygenic risk score (PRS) for head and neck cancer overall and separately for oral cavity cancer and oropharyngeal caner. The PRS was estimated as the sum of the number of risk alleles one carries weighted by the log odds ratio derived from the GWAS studies reported to date except for 8 variants in the HLA region that were identified using HPV-positive oropharyngeal cancer cases³⁰, where the weights were calculated based on the present study participants.

Exposure variables and cancer endpoints

The demographic and lifestyle factors to be considered in the prediction model were defined a priori based on the previous literature. These factors included age, tobacco smoking history (smoking status and pack-years), alcohol consumption history (drinking status and amount of alcohol consumed) and education (postsecondary education as reference). Body mass index was not included in the model, because it was mostly collected at the time of cancer diagnosis and might have been influenced by disease occurrence or progression. In addition to the above predictors, we also included HPV serostatus in the model for oropharyngeal cancer, and PRS in the model for oral cavity cancer and oropharyngeal cancer. Since a vast majority of the study population (91.4%) self-identified as European descendants, we limited our analysis to those with European ancestry.

All cancer cases were coded according to the International Classification of Disease Volume 10 (ICD-10). In the present analysis, cancer cases were classified as 1) oral cavity cancer: cancers of the lip (C00.3-C00.9, C02.0-C02.3), gum (C03.0, C03.1, C03.9), floor of mouth (C04.0, C04.1, C04.8, C04.9, C05.0) and other and unspecified parts of mouth (C06.0, C06.1, C06.2, C06.8, C06.9); 2) oropharyngeal cancers: cancers of the base of tongue/lingual tonsil (C01.9, C02.4), soft palate (C05.1), uvula (C05.2), palatine tonsil (C09.0, C09.1, C09.8, C09.9) and oropharynx (C10.0, C10.2-C10.9); and 3) other head and neck cancer: cancers of the salivary gland (C07.9C08.9), nasopharynx (C11.0-C11.9), hypopharynx (C12.9-C13.9), oral cavity-oropharynxhypopharynx not otherwise specified (C02.8, C02.9, C05.8, C05.9, C14.0, C14.2, C14.8) and larynx (C10.1, C32.0-C32.9). Head and neck cancer included cancers of all the above sites.

Model development and evaluation

Given the substantially different incidence of head and neck cancers by sex, we developed and evaluated the risk model separately for men and women from the outset. For the purpose of model development and evaluation, we randomly divided the data in each study into 70% training set for model development and 30% hold-out testing set for model performance evaluation (Supplementary Figure 1). In the training set, we included all statistically significant variables from the univariate logistic regression of the putative risk factors and performed backward stepwise selection to determine the final panel of variables. The linearity was visually inspected by plotting continuous variable against the logit of the outcome and the Box-Tidwell test. Those variables that appear to show a nonlinear relationship were modeled as categorical variables in subsequent analyses. Interactions between variables were evaluated by including product terms of the risk factors in the model. Missing values for lifestyle and demographic variables were imputed using multiple imputation- we created ten imputed datasets by chained equations procedure in which all predictor variables were used to impute missing values. Models were then fitted to each imputed dataset and the results were pooled using Rubin’s rule³⁵. Since only two studies had information on family history of head and neck cancer, multiple imputation was not performed for this variable. For variables with multiple measures (such as cigarette and alcohol use status and intensity), we selected a variable based on the Akaike information criterion. The models’ ability to discriminate was assessed through Area under the Receiver Operating Characteristic Curves (AUC) in the hold-out testing set. To evaluate the model calibration prospectively on the absolute risk scale, we used the UK Biobank data with longitudinal follow-up (Supplemental methods). The model calibration was evaluated by calibration plot comparing the predicted versus the observed probability (defined as empirical proportion of the outcome), and Hosmer-Lemeshow goodness-of-fit test.

Estimation of absolute risk

The five-year absolute risk of developing head and neck cancer was estimated based on Cox proportional hazards model, accounting for age-specific competing hazards of mortality of other causes. The absolute risk within a given time interval was estimated by integrating (i) a model of relative risks, (ii) age-specific incidence of head and neck, oral or oropharyngeal cancer, and (iii) distribution of the risk factors in the population of interest (Supplementary Methods). The details of methods have been described in detail previously^{36, 37}. The distribution of risk factors was approximated using the UK Biobank population cohort^{38, 39}. The age-specific cancer rates and competing hazards for mortality (Supplementary Table 2) were obtained from Surveillance, Epidemiology, and End Results (SEER) Program⁴⁰ and Centers for Disease Control and Prevention, National Center for Health Statistics database⁴¹ respectively. Since the effect of smoking and alcohol drinking on oropharyngeal cancer may differ by HPV serostatus, we estimated the effect of these risk factors stratified by HPV serostatus, and use the stratum-specific effect estimates for the absolute risk trajectory. A standard non-parametric bootstrap method was used to compute 95% confidence bands of the absolute risk estimates corresponding to the highest risk stratum. Relative risks were estimated from the bootstrap re-samples of the multiple-imputed model building dataset, while age-specific incidence rates, competing mortality rates and the reference dataset were kept constant. All analyses were performed in R statistical software (version 4.0.3): the mice and psfmi package for multiple imputation and pooling and iCARE package for absolute risk estimation.

Results

The distribution of key characteristics of all study participants are shown in Table 1. The study population included more males than females. Cancer cases had higher smoking prevalence and greater pack-years history compared to controls. The average alcohol consumption amount was also higher in cases. As expected, the proportion of participants with HPV seropositivity was much higher among cancer cases, specifically among patients with oropharyngeal cancer. The distributions of risk factor in hypopharynx cancer and larynx cancer showed similar patterns to that of head and neck cancer (Supplementary Table 3).

Table 1.

The key characteristics of the study populations

Variables	Categories	Head and neck cancer	Oral cavity cancer	Oropharyngeal cancer	Controls
Total (n)		10126	2431	3727	5254
Study (n)
	CHANCE	1010	158	277	1114
	ARCAGE	1924	470	411	2043
	PITTSBURGH	847	263	365	811
	TORONTO	1663	400	790	1286
	HN5000	4682	1140	1884	-
Sex, n (%)
	Men	7750 (76.6)	1572 (64.7)	2976 (79.8)	3471 (66.1)
	Women	2373 (23.4)	859 (35.3)	751 (20.2)	1783 (33.9)
	Missing	3	0	0	0
Age (years), mean (SD)		60.7 (10.9)	61.8 (12.0)	58.7 (9.3)	60 (12.0)
Tobacco Smoking status, n (%)
	Never	1638 (18.9)	448 (21.5)	780 (24.9)	2135 (40.8)
	Former	3510 (40.5)	744 (35.6)	1346 (43.0)	2009 (38.4)
	Current	3527 (40.7)	895 (42.9)	1006 (32.1)	1094 (20.9)
	Missing	1447	342	592	16
Tobacco Pack-years, median (IQR)		36 (33.0)	34.1 (32.5)	30.0 (32.0)	21.0 (30.5)
Alcohol drinking status, n (%)
	Never	1797 (20.7)	484 (23.2)	665 (21.0)	1080 (20.6)
	Former	4351 (50.1)	1015 (48.6)	1793 (56.7)	1493 (28.5)
	Current	2531 (29.2)	588 (28.2)	702 (22.2)	2667 (50.9)
	Missing	1443	343	563	14
Drink/week, median (IQR)		20.3 (29.0)	20.5 (29.0)	17.9 (28.0)	7.4 (12.6)
Education, n (%)
	Postsecondary	2377 (29.8)	523 (27.1)	1043 (35.3)	2585 (52.1)
	High school diploma	2419 (30.4)	622 (32.2)	997 (33.7)	1030 (20.8)
	None/elementary	3168 (39.8)	785 (40.7)	918 (31.0)	1345 (27.1)
	Missing	2163	498	766	294
HPV serostatus, n (%)a
	Total tested			1804	2332
	Negative			660 (36.6)	2312 (99.1)
	Positive			1144 (63.4)	20 (0.9)
Polygenic risk score, median (IQR)	Total genotyped	3901	1339	1823	2962
		0.47 (0.05–0.82)	−0.002 (−0.21 – 0.16)	0.21 (−0.26 – 0.62)	0.24 (−0.26 – 0.66)

^aHPV serology status is defined based on high HPV16 E6 antibody levels (>1000 median fluorescence intensity, MFI) or seropositivity for three of four HPV16 early proteins (E1: >200 MFI, E2: >679 MFI, E6: >484 MFI and E7: >548 MFI).

Pack-years and alcohol intensity both showed non-linear association with cancer risk; thus, they were modelled as categorical variables in subsequent analysis. Smoking pack-years was categorised into never, moderate and heavy smokers with the cut-off for the latter two categories being the sex-specific median value of ever smokers among controls (Supplementary Table 4). Drinking intensity and PRS were divided into sex-specific tertiles based on the distribution among controls (Supplementary Table 4). We did not detect significant interaction between variables and are not included in the final model (Supplementary Table 5a and 5b). Table 2 shows the odds ratios and 95% confidence interval for developing head and neck cancers in the final multivariable model by sex. Overall, in both men and women, smoking, heavy alcohol drinking, lower education, HPV seropositivity, and higher PRS were positively associated with head and neck cancer risk. The association of these factors with oropharyngeal cancer and oral cavity cancer showed similar patterns, albeit the magnitude of the risk estimate was greater for oropharyngeal cancer for smoking and drinking (Table 2).

Table 2.

Odds ratios (ORs) and 95% confidence intervals (CIs) for developing head and neck cancers and key risk factors by sex based on multivariable logistic regression models

Variable	Categories	Head and neck cancer	Oral cavity cancer	Oropharyngeal cancer
		OR (95%CI)*	OR (95%CI)*	OR (95%CI)*
Men
Smoking statusa
	Never	1 (Ref.)	1 (Ref.)	1 (Ref.)
	Moderate	1.22 (1.02–1.44)	1.53 (1.14–2.05)	1.48 (1.00–2.20)
	Heavy	2.52 (2.11–3.01)	3.15 (2.45–4.06)	5.14 (3.54–7.47)
Drinking statusb
	Never/Low	1 (Ref.)	1 (Ref.)	1 (Ref.)
	Moderate	0.87 (0.74–1.03)	0.88 (0.69–1.13)	0.83 (0.57–1.20)
	Heavy	1.66 (1.41–1.94)	1.82 (1.46–2.27)	2.27 (1.69–3.04)
Education
	Postsecondary	1 (Ref.)	1 (Ref.)	1 (Ref.)
	High school diploma	2.26 (1.93–2.65)	2.74 (2.14–3.51)	1.93 (1.4–2.65)
	None/elementary	1.96 (1.65–2.34)	2.71 (2.09–3.51)	2.51 (1.85–3.4)
HPV serostatus
	Negative			1 (Ref.)
	Positive			385 (218–681)
Polygenic risk scorec
	Low (1st tertile)	1 (Ref.)	1 (Ref.)	1 (Ref.)
	Middle (2nd tertile)	1.52 (1.29–1.79)	1.33 (1.05–1.69)	1.14 (0.85–1.55)
	High (3rd tertile)	2.35 (2.01–2.75)	2.16 (1.72–2.71)	1.59 (1.19–2.13)
Women
Smoking statusa
	Never	1 (Ref.)	1 (Ref.)	1 (Ref.)
	Moderate	1.32 (1.01–1.71)	1.37 (0.96–1.95)	2.09 (1.09–4.00)
	Heavy	3.46 (2.75–4.35)	3.23 (2.41–4.32)	6.86 (4.14–11.36)
Drinking statusb
	Never/low	1 (Ref.)	1 (Ref.)	1 (Ref.)
	Moderate	0.70 (0.53–0.92)	0.73 (0.51–1.06)	1.12 (0.64–1.97)
	Heavy	1.49 (1.18–1.89)	1.50 (1.11–2.03)	2.64 (1.66–4.18)
Education
	Postsecondary	1 (Ref.)	1 (Ref.)	1 (Ref.)
	High school diploma	3.17 (2.52–4.00)	3.81 (2.80–5.18)	3.56 (2.17–5.84)
	None/elementary	4.88 (3.77–6.32)	6.44 (4.69–8.84)	5.04 (2.99–8.50)
HPV serostatus
	Negative			1 (Ref.)
	Positive			237 (103–550)
Polygenic risk scorec	Low (1st tertile)	1 (Ref.)	1 (Ref.)	1 (Ref.)
	Median (2nd tertile)	1.71 (1.33–2.19)	1.71 (1.24–2.37)	0.90 (0.54–1.52)
	High (2nd tertile)	1.89 (1.48–2.42)	2.07 (1.51–2.84)	1.25 (0.77–2.01)

^*The odds ratio estimates are based on all factors included in this table in the multivariable model.

^aThe cut-off is based on sex-specific medians among ever smokers in the control group: <24 pack-years (Moderate smoker) or ≥24 pack-years of smoking (Heavy smoker) in men; and <14 pack-years (Moderate smoker) or ≥14 pack-years of smoking (Heavy smoker) in women.

^bThe cut-off is based on sex-specific tertiles in the control group; <5.5 drinks/week (Never/low drinker), 5.5 to <14.7 drinks/week (Moderate drinker) or ≥14.7 drinks/week (Heavy drinker) in men; and <2.2 drinks/week (Never/low drinker), 2.2 to <6.9 drinks/week (Moderate drinker) or ≥6.9 drinks/week (Heavy drinker) in women.

^cThe polygenic risk scores are computed for oral cavity and oropharyngeal cancer separately based on the loci reported for these tumor types. Loci reported for head and neck cancer or their anatomical subsites are included in the PRS for head and neck cancer overall.

To assess whether the inclusion of a case-only cohort (HN5000) affected our results, we conducted a sensitivity analysis by excluding all HN5000 cases. There was little to no meaningful change of any estimates of the factors included in the model (Supplementary Table 6) when including HN5000, thus our primary analysis was based on the full dataset, from which the estimates have higher precision.

The predictive performance of the models in the hold-out testing set based on epidemiological risk factors and the addition of HPV serostatus and PRS is shown in Table 3 and Supplementary Figure 2. In men, the addition of PRS to the model with only epidemiological risk factors improved the discriminative accuracy of the model from AUC of 0.69 to 0.72 (95% CI=0.69–0.75) for head and neck cancer overall, and to 0.73 (95% CI=0.69–0.77) for oral cavity cancer. In women, adding PRS only improved the predictive accuracy for oral cavity cancer, but not for head and neck cancer overall with the resulting AUCs of 0.79 (95% CI=0.74–0.83) and 0.75 (95% CI, 0.71–0.79) respectively. For oropharyngeal cancer, addition of HPV serostatus to the model with only epidemiological risk factor greatly improved the predictive accuracy of the model in both men and women, resulting in the AUCs of 0.92 (95% CI, 0.90–0.94) and 0.91 (95% CI, 0.86–0.94), respectively. Further addition of the PRS marginally improved the predictive accuracy, with AUC of 0.94 (95% CI, 0.92–0.95) in men, and with AUC of 0.92 (95% CI, 0.88–0.95) in women. Assessment of the predictive performance of the models by 10-year age categories showed comparable AUCs in each age strata to the overall AUC for all three cancer types, with small variations, albeit wider confidence intervals. For example, the AUCs of the full model for OPC in women were 0.94 (95%CI, 0.87–0.97), 0.90 (95%CI, 0.81–0.95), 0.91 (95%CI, 0.76–0.97) and 0.91 (95%CI, 0.73–0.98) for age strata of less than 55 years old, 55–64, 65–74 and 75 years and older, respectively.

Table 3.

Area Under the Receiver Operating Characteristic Curves (AUCs) of risk prediction models for head and neck cancer in hold-out testing set

Model	Men, AUC (95%CI)			Women, AUC (95%CI)
	Head and neck cancer	Oral cavity cancer	Oropharyngeal cancer	Head and neck cancer	Oral cavity cancer	Oropharyngeal cancer
Epidemiological risk factors	0.69 (0.67–0.71)	0.69 (0.66–0.72)	0.66 (0.64–0.69)	0.75 (0.72–0.78)	0.75 (0.71–0.79)	0.76 (0.72–0.80)
Epidemiological risk factors and HPV serostatus			0.92 (0.90–0.94)			0.91 (0.86–0.94)
Epidemiological risk factors and PRS	0.72 (0.69–0.75)	0.73 (0.69–0.77)	0.71 (0.67–0.74)	0.75 (0.71–0.79)	0.79 (0.74–0.83)	0.76 (0.71–0.81)
Epidemiological risk factors, HPV serostatus and PRS			0.94 (0.92–0.95)			0.92 (0.88–0.95)

Epidemiological risk factor model includes age, smoking packyears, alcohol drinking intensity and education.

HPV, human papillomavirus; PRS, polygenic risk scores.

As a secondary sensitivity analysis, we tested the model performance based on HPV serostatus defined by HPV16 E6 antibody levels (> 1000 MFI) alone to assess the potential loss in predictive accuracy for oropharyngeal cancer. It showed similar AUCs to that of models containing HPV status defined by multiple markers. When HPV seropositivity was defined by HPV16 E6 alone, the AUC for the full model was 0.93 (95%CI=0.90–0.94) in men and 0.89 (95% CI, 0.84–0.93) in women.

Finally, we estimated 5-year absolute risk of head and neck cancer according to risk factor profiles including all aforementioned risk factors included in the final model using the UK Biobank population cohort. The model calibration is shown in Supplementary Figure 3. In general, the models are well calibrated based on calibration slope close to 1 and the Hosmer-Lemeshow test did not indicate deviation for most of the models, except for oropharyngeal cancer in men which has limited sample size and therefore is subject to fluctuations.

Figure 1 shows the absolute risk estimates for overall head and neck cancers. As expected, the absolute risk estimate increased with older age. In general, the risk was low among never users of cigarettes or alcohol in both men and women whereas the risk increased with the heavy use of these substances. The estimated 5-year absolute risk among heavy smokers and heavy drinkers at age 65 varied from 0.64% in the lowest PRS tertile to 1.20% in the highest PRS tertile in men and from 0.23% to 0.30% in women. Since risk profiles are different by anatomical site, we also estimated the 5-year absolute risk separately for oral cavity (Figure 2) and oropharyngeal cancer (Figure 3). For oral cavity cancer, smoking and drinking accounted for substantial variation in the risk conferred, with those heavy users of both tobacco and alcohol being the highest risk group. In general, the 5-year risk was higher among those with higher PRS in both sexes, but remained low in general (Figure 2).

Figure 1. Five-year absolute risk estimates of head and neck cancer stratified by tobacco smoking, alcohol drinking and polygenic risk score for men and women.

The blue, yellow and red lines represent never, moderate and heavy tobacco smokers, respectively. The dashed and solid lines represent never/moderate and heavy alcohol drinkers. For example, yellow solid line represents moderate smokers who drank heavily. The gray zone represents the 95% confidence intervals of the highest risk category. The smoking category (Moderate vs Heavy) cut-off is based on sex-specific medians among ever smokers in the control group. The alcohol drinking categories (Low, Moderate, Heavy) and polygenic risk score (Low, Medium and High) are based on sex-specific tertiles in the control group (Supplementary Table 4).

Figure 2. Five-year absolute risk estimates of oral cavity cancer stratified by smoking, drinking and polygenic risk score for men and women.

The blue, yellow and red lines represent never, moderate and heavy tobacco smokers, respectively. The dashed and solid lines represent never/moderate and heavy alcohol drinkers. For example, yellow solid line represents moderate smokers who drank heavily. The gray zone represents the 95% confidence intervals of the highest risk category. The smoking category (Moderate vs Heavy) cut-off is based on sex-specific medians among ever smokers in the control group. The alcohol drinking categories (Low, Moderate, Heavy) and polygenic risk score (Low, Medium and High) are based on sex-specific tertiles in the control group (Supplementary Table 4).

Figure 3. Absolute risk estimates of oropharyngeal cancer stratified by tobacco smoking, alcohol drinking and human papillomavirus (HPV) serostatus for men and women.

The color of the lines represents different smoking and drinking categories. The solid and dashed line represent HPV seropositive and seronegative, respectively. The dotted line represents the average risk among HPV seropositive individuals, irrespective of their tobacco and alcohol consumption status. The smoking category (Moderate vs Heavy) cut-off is based on sex-specific medians among ever smokers in the control group. The alcohol drinking categories (Low, Moderate, Heavy) and polygenic risk score (Low, Medium and High) are based on sex-specific tertiles in the control group (Supplementary Table 4).

On the other hand, we observed a substantial range of 5-year risk for oropharyngeal cancer and HPV seropositivity status accounted for the majority of the risk variation (Figure 3). While the 5-year risk remained very low among those who are HPV-seronegative (<0.1%), the 5-year risk of those who are HPV-seropositives are considerably higher. For example, irrespective of the tobacco and alcohol consumption, the average risk of developing oropharyngeal cancer among HPV seropostives of a 60-year old was 8.1% for men and 2.2% for women (Figure 3). In addition, there are differential risk trajectories based on individual’s risk profiles. For example, the average 5-year risk for a 60-year old man, HPV-seropositive, lifetime non-drinker and non-smoker was 5.8%, and it increased up to 14.9% for heavy smokers and heavy drinkers, with the other parameters being held constant, albeit wide confidence limits. The corresponding risk estimates for a 60-year old HPV seropositive, lifetime non-drinker and non-smoker woman was 1.3% and in HPV seropositive, heavy smokers and heavy drinkers it was 4.4% (Figure 3). For oropharyngeal cancer, due to the very small number of HPV seropositive observations in our control population, we could not estimate the absolute risk by PRS, in conjunction with HPV serostatus.

Discussion

To our knowledge, this is the first study to develop a prediction model for head and neck cancer using HPV serostatus and genetic factors along with known or potential risk factors in European-descent population. The inclusion of HPV serostatus along with epidemiological risk factors improved the model’s predictive performance for oropharyngeal cancer. By integrating a US national database of incidence and mortality rates, we observed diverse trajectories by risk factor profiles including HPV serostatus and PRS after accounting for competing risks. Those with HPV seropositive reached high risk level for OPC that could benefit from primary prevention strategy or intensive surveillance, which is currently lacking. These results suggest that risk prediction models can be useful in identifying the population at higher risk of developing head and neck cancer, with the risk varying by anatomical sites and individual risk profiles.

Demographic and lifestyle factors including age, cigarette smoking, alcohol drinking and education were found to be significant predictors of the head and neck cancer risk in our model. The predictive accuracy of our model for oropharyngeal cancer was over 90% when including HPV serostatus, which represents improvements from previous prediction models^20–23. Given that HPV occurrence is rare for oral cavity cancer, we did not include HPV serostatus as a predictor in the model for oral cavity cancer. However, inclusion of PRS showed modest improved performance for oral cavity cancer, suggesting that combination of multiple risk loci may provide value in oral cancer risk prediction.

HPV16 E6 antibodies are considered to be markers of risk of oropharyngeal cancer. In an analysis using prospectively collected plasma samples from a cohort of European subjects¹⁶, HPV16 E6 seropositivity was associated with a more than 100-fold increase in risk of oropharyngeal cancer¹⁶. More importantly, this association remained strong based on samples collected more than 10 years before diagnosis¹⁶. This suggests that HPV16 E6 antibody may have utility as a biomarker for risk stratification of developing oropharyngeal cancer prior to cancer diagnosis. However, the long lead time between HPV seropositivity and cancer diagnosis could pose challenges in screening implementation, with respect to the timing and frequency of screening and potential psychological burdens due to years of continuous evaluation^{13, 42}. On the other hand, the challenges posed by the long lead time of HPV serological markers are not completely distinct from other non-modifiable risk factors such as demographics or genetic susceptibility, which highlights the importance of estimating the absolute risks within a specific time interval using age as the time horizon to determine the optimal time point of actionability, which is the focus of the present study.

In general, screening efficacy depends on pre-cancerous lesions that can be identified with high sensitivity and specificity. Currently, no screening guidelines exist for the early detection of head and neck precancerous lesions or cancers in the general population. For oropharyngeal cancer, while the risk level for the majority of the population is too low to warrant population-based screening, we did observe that the absolute risk trajectory varied greatly by individual’s risk factor profiles including smoking, drinking and HPV serostatus. The differentiation of risk trajectory among HPV-seronegatives and HPV-seropositives was predominately depending on the consumption of tobacco and alcohol. We showed that HPV seropositive status led to a high predictive performance, which raises the potential of HPV serology-based test for screening oropharyngeal cancer. In our study, although we used a compound definition of multiple HPV serologic markers, HPV16 E6 seropositivity was the primary driving determinant that defined HPV seropositivity in the majority of participants. Our sensitivity analysis showed that there is limited loss in the predictive accuracy when using HPV16 E6 alone. This suggests that HPV16 E6 antibody is an adequate test to determine the seropositivity, which may help to improve the feasibility of large-scale population testing.

However, the main challenges remain that pre-cancer lesions for oropharyngeal caner have not been identified⁴², and given the relatively low incidence of oropharyngeal cancer, the HPV serology-based test would result in low positive predictive value. Both of these factors would limit the balance between psychologic and physical distress related to screening and the potential benefits^{13, 43}. Given the low prevalence of HPV16 early protein antibodies in the general population, further studies are required to evaluate the effectiveness of screening modalities in secondary prevention of oropharyngeal cancer, as well as the risk-threshold to maximize the cost-efficiency, which is beyond the scope of the present work. Nonetheless, in terms of primary prevention, our model may be informative for individuals at high-risk and potentially encouraging behavioral modifications, such as intensive smoking cessation programs.

Regarding oral cancer, the US Preventive Services Task Force concluded that the current evidence was insufficient to assess the balance of benefits and harms of screening for oral cancer in asymptomatic adults⁴⁴. A large trial conducted in India, where participants were randomly assigned to receive visual screening (of the oral cavity by trained healthcare workers every three years for four rounds) versus the usual care (control group), reported reduced mortality from oral cancers in the screened group which was mainly observed in tobacco and alcohol users⁴⁵. In another report of a nationwide, population-based screening program for oral cancer in Taiwan, the mortality of oral cancer was reduced by 50% in the screening group compared to the expected oral cancer mortality in the absence of screening⁴⁶. These studies suggest benefit of screening for oral cancer in high-risk groups. However, these studies were conducted in populations with higher incidence of oral cancers and may not be directly generalizable to other populations with different risk profiles. Although other visual adjunctive technologies such as toluidine blue, brush biopsy or fluorescence imaging have been evaluated for oral cancer screening, their effectiveness as a screening tool to reduce oral cancer mortality is not established^{47, 48}.

Our study has several limitations. First, the study participants represented a population of European ancestry and thus the model may not be generalizable to other ethnicities with different risk factor profiles. Nonetheless, in comparison to the large national survey⁴⁹, we found that the risk profiles, mainly cigarette smoking and alcohol drinking in our VOYAGER study is comparable with the large national survey data: 19.3% of the population were current smoker in the survey versus 20.9% in our study, and 78.9% of the population were ever drinkers in the survey population compared to 79.4 in our study. If there was bias introduced due to the source data for risk factor distribution, it would likely to be minimal. On the other hand, the absolute risk was estimated based on UK Biobank population cohort, which has been recognized as a healthier cohort⁵⁰. Therefore, the estimated absolute risks maybe lower than in the general population⁵⁰. Second, even though there was large number of cases for overall head and neck cancer analysis, the sample size was small for analysis by subsite, particularly with HPV serostatus and genetic data. Cautious interpretation of the extreme high-risk group is needed given the wide confidence bands, in particular for oropharyngeal cancer. Third, only a subset of study participants had information on family history of head and neck cancer in our study and thus this variable was not included in the model. Fourth, our data contributed to the original discovery of the susceptibility loci of head and neck cancers, therefore the PRS effect may be overfitted. Future studies should use independent datasets to reduce the possibility of overfitting in the PRS model. Finally, we were not able to conduct external validation of our model given the limited data availability that include both HPV serology and genetic data available outside of the current participating studies.

In summary, we developed the first absolute risk prediction model for head and neck cancer which incorporated all key aspects including environmental risk factors, HPV serostatus, and genetic risk variants. The model performance was improved compared to previous models based on epidemiologic factors only, and it may be useful for stratifying populations at high risk of developing head and neck cancer. Future validation of these models based on prospective cohorts would be warranted. Nonetheless, the high absolute risk level among those with HPV seropositive highlights the need to consider primary prevention and intensive surveillance for OPC in targeted subgroup.

What’s new?

Based on 5 large datasets, this is the first study of an integrated head and neck cancer risk model, including lifestyle risk factors, polygenic risk score, and human papillomavirus serology specifically for oropharyngeal cancer. The models are well-calibrated and showed excellent predictive accuracy. To determine the translational value of these models, we estimated the head and neck cancer absolute risk within the next 5 years using age as the time horizon to determine the optimal time point of actionability. Specifically for oropharyngeal cancer, it showed a distinctive absolute risk trajectory of approximately 3-fold difference for both men and women by risk profiles, with the average risk among human papillomavirus seropositive reaching to 8.1% in men and 2.2% in women at age 60. These risk levels indicate the need of primary prevention or intensive surveillance for the targeted subgroup which is currently lacking.

Abbreviations:

AUC

area under the receiver operating characteristic curves

CI

confidence interval

HPV

human papillomavirus

MFI

median fluorescence intensity

OR

odds ratio

PRS

polygenic risk score

Data Availability Statement

Data sources and handling of the publicly available datasets used in this study are described in the Materials and Methods. Further details and other data that support the findings of this study are available from the corresponding authors upon request.