Revisit of Field Cancerization in Squamous Cell Carcinoma of Upper Aerodigestive Tract: Better Risk Assessment with Epigenetic Markers

Squamous cell carcinoma (SCC) is the major phenotype of esophageal cancer in Asia. Survival of this cancer is decreased because of field cancerization, in which synchronous and metachronous cancers occur in the upper aerodigestive tract (UADT; ref. 1). Screening and surveillance measures, particularly the use of sophisticated endoscopic technologies such as narrow-band imaging and Lugol staining (2), currently aid physicians in starting early intervention and providing curative treatment.

The populations of patients exposed to well-established risk factors may outnumber the capacity of endoscopists to screen them. For instance, among the adult male population in Taiwan, 46% drink alcohol, 15% chew betel quid, and 57% smoke cigarettes (3), which are all behaviors that result in an increased incidence of SCC (4). Following a negative endoscopy, an efficient risk-assessment method is needed to design surveillance programs for early detection of superficial SCC in the head and neck region and esophagus (2). However, past efforts based on demographic characteristics (5–7), lifestyle risk factors (8, 9), genetic susceptibilities (10–14), serological markers (15–18), and endoscopic findings (19), alone or in combination (15, 20, 21), have not produced models efficient enough to be effective.

Aberrant DNA methylation in histologically normal mucosae has attracted attention as an indicator of past exposure to carcinogens and as a marker for future risk prediction (22). Reliable measuring techniques, such as real-time quantitative methylation-specific PCR (qMSP), have been confirmed to be sensitive and accurate in quantifying this early carcinogenic event (23). They offer an opportunity to elucidate whether aberrant DNA methylation can serve as a marker to quantify the field effects associated with SCC of the UADT. If this hypothesis is found to be true, we can compare the performance of risk-assessment methods based on epigenetic markers with those based on traditional risk factors. We also validated the performance of the epigenetic markers with a prospectively recruited independent cohort.

We recruited consecutive Taiwanese patients who received endoscopic screening at the National Taiwan University Hospital between January 2008 and May 2011 due to UADT symptoms, such as globus sensation and dysphagia. This cohort was split into a training set to develop a risk-assessment model for UADT cancer based on epigenetic markers and a validation set to evaluate model performance. Our initial pilot experiment determined the required sample size for hypothesis testing with 123 biopsied samples from 103 patients. The study protocol included thoroughly evaluating traditional risk factors, screening by endoscopy, and quantifying DNA methylation in the biopsied specimens. All participants provided informed consent, and the Ethics Committee of the National Taiwan University Hospital approved the study (no: 200706034R and 200806039R).

Prior to the endoscopic screening, patients underwent face-to-face interviews, physical examinations, and laboratory tests to identify traditional risk factors for SCCs in the head and neck region (24) and in the esophagus (25) that would be used as baseline comparators for the upcoming epigenetic approach. These included: the demographic characteristics (6, 7) of age, sex, and body mass index (BMI); the lifestyle risk factors (8, 9, 11) of alcohol drinking, betel quid chewing, and cigarette smoking (briefly as “ABC”); polymorphisms in genes encoding enzymes involved in the metabolism of alcohol (10–13), including aldehyde dehydrogenase (ALDH) 2, alcohol dehydrogenase (ADH) 1B, and ADH 1C; polymorphisms in genes encoding enzymes involved in the metabolism of xenobiotics (14), including glutathione S transferase (GST) P1, GST M1, and GST T1; and serological markers (15–18), including increased mean corpuscular volume (MCV), Helicobacter pylori infection, and human papillomavirus infection.

Lifestyle risk factors were recorded according to frequency (alcohol where one time indicates at least 15.75 gm of ethanol: 0, never; 1, once per week; 2, once to twice per week; 3, 3–4 times per week; and 4, 5 or more times per week; betel quid: 0, never; 1, less than 1 piece daily; 2, 1–10 pieces daily; 3, 11 to 20 pieces daily; and 4, more than 20 pieces daily; cigarettes: 0, never; 1, less than 0.5 pack daily; 2, 0.5–1 pack daily; 3, 1–2 packs daily; and 4, more than 2 packs daily) and duration (0, never; 1, less than 1 year; 2, 1–10 years; 3, 11–20 years; and 4, more than 20 years). The cumulative lifetime exposure was calculated by multiplying the frequency and duration, and then categorized into 3 levels: level 1, never (0); level 2, light to moderate (1–11); and level 3, heavy (≥12). Polymorphic genotypes were determined from peripheral blood leucocytes with PCR-RFLP (the primers are provided in the Supplementary Table S1). Diagnosis of H. pylori infection was made histologically or serologically (Diagnostic Products Co.; ref. 26) as appropriate. Human papillomavirus infection was serologically diagnosed by ELISA with commercially available kits (Wuhan EIAab Science Co., Ltd.; ref. 27). The cutoff value for human papillomavirus IgG was set at 600 U/mL based on 3 SDs above the mean optical density obtained from serum samples of negative controls.

After evaluating traditional risk factors, patients underwent endoscopic examinations by white-light, narrow-band imaging, and Lugol staining in a single session, which is a method shown to be accurate in the detection of cancers of the UADT (28, 29). We evaluated endoscopic findings of numerous irregular-shaped multiform unstained areas over the background esophageal mucosae, namely, Lugol voiding lesions (LVL) after spraying of Lugol solution (19), as one of the traditional risk factors. A biopsy was done on any suspicious lesions along the UADT before chemotherapy or radiation therapy and also on normal-appearing mucosae in the esophagus. Normal-appearing mucosa was defined as an area that did not seem brownish (“brownish” indicated microvascular proliferation) under narrow-band imaging and, after being sprayed with Lugol's solution, was stained (“unstained” indicated glycogen depletion). For patients with either head and neck cancer or esophageal cancer, biopsies for normal-appearing mucosae were routinely done 3 cm above the squamocolumnar junction in the esophagus; in cases with neoplastic lesions close to this area, the biopsy sites were moved to at least 5 cm away from the neoplastic margin. In addition to histologic confirmation, a set of specimens was stored at −80°C until genomic DNA was extracted.

Genomic DNA was sent to the National Cancer Center Research Institute in Japan. The DNA methylation of the epigenetic markers was quantified with the qMSP. The laboratory work was similar to the details previously reported that used an independent test set of 60 tumor samples derived from Japanese individuals (30). In brief, 14 genes whose promoter CpG islands were methylated in esophageal SCCs were isolated by a genome-wide screening of genes reexpressed after esophageal SCC cell lines were treated with the demethylating agent 5-aza-2′-deoxycytidine. Most of the 14 genes were also methylated at low levels in adjacent esophageal mucosae. We analyzed 4 genes correlated with smoking duration in this study: homeobox A9 (HOXA9), neurofilament heavy polypeptide (NEFH, 200 kDa), ubiquitin carboxyl-terminal esterase L1 (UCHL1), and metallothionein 1M (MT1M).

Laboratory analyses were done by laboratory technicians who were blinded to the clinical status of the samples analyzed. As in our previous report (30), DNA was digested by BamHI, denatured, and treated by 15 cycles denaturing for 30 seconds at 95°C and incubating for 15 minutes at 50°C in 3.1 N sodium bisulfite (pH 5.0) and 0.5 mmol/L hydroquinone. The samples were desalted and then desulfonated in 0.3 N NaOH. The qMSP was done with primer sets specific to either the methylated or unmethylated sequence (M or U set) with the sodium bisulfite-treated DNA and an iCycler Thermal Cycler (Bio-Rad Laboratories). We compared the amplification of sample DNA to that of standard DNA that was prepared from a plasmid with a cloned PCR product and that had known copy numbers of the cloned DNA. The amount of methylated and unmethylated DNA in the sample was quantified.

Data are expressed as average ± SD (median; range) for continuous variables and as a percentage for categorical variables. The measurement of DNA methylation levels in the biopsied specimens was the starting point for the following statistical analyses: (i) comparing methylation levels according to stage of carcinogenic sequence and level of carcinogen exposure; (ii) building logistic-regression models to determine the likelihood of UADT cancers based on epigenetic markers, traditional risk factors, and both together; and (iii) comparing the predictive value of epigenetic markers on a validation set of patients not used in the initial model building phase.

First, the ability of each epigenetic marker to serve as an indicator for the field effects was confirmed by seeking a stepwise relationship among the methylation levels that followed the carcinogenic sequence related to the chronic irritation of ABC, which are the 3 principal determinants of SCC. We simulated the carcinogenic sequence by categorizing the biopsied specimens into 4 groups: group 1, normal esophageal mucosae of healthy subjects with no exposure to ABC; group 2, normal esophageal mucosae of healthy subjects with at least one of the above exposures; group 3, normal esophageal mucosae of cancer patients; and group 4, cancerous mucosae. Methylation levels were expected to increase in a stepwise manner in linear regression analyses as the group number increased. Scatter plots were drawn to describe the results. Pairwise comparisons between groups were also made using the nonparametric Mann–Whitney U test. We also correlated methylation levels in the normal-appearing esophageal mucosae (i.e., groups 1, 2, and 3) with cumulative lifetime exposures to ABC. Tests for linear trends and pairwise comparisons between subjects with different exposure levels were done in a similar manner. In cancerous tissues (i.e., group 4), the association between methylation levels and clinicopathologic characteristics was evaluated.

Second, the predictive value of our epigenetic markers was assessed by developing logistic-regression models that used the methylation levels in normal-appearing esophageal mucosae to predict the probability of a patient having head and neck cancer, esophageal cancer, and UADT cancer based on the training samples. Following univariate logistic-regression analyses, an epigenetic model was built using backward elimination of predictors from the full 4-predictor model that excluded those with P ≥ 0.05. We excluded strong linear dependencies among the predictors by examining tolerance and the variance inflation factor. We also excluded the statistical significance of interactions and nonlinear terms.

We developed logistic-regression models using the traditional risk factors to compare with the epigenetic model. Continuous predictors included age, BMI, levels of cumulative lifetime exposures to ABC, and MCV. For the categorical predictors, the baseline comparators were: female, unexposed to ABC, without H. pylori infection, without human papillomavirus infection, and carrying these genotypes: ALDH2*1/2*1 (fast metabolizers), ADH1B*1/*2 or *2/*2 (fast), ADH1C*1/*2 or *2/*2 (slow), GSTP1 ile/val or val/val (fast), GSTM1 present (active), and GSTT1 present (active). Univariate logistic-regression analyzed each predictor and those predictors with P < 0.05 were included in a multivariate regression model. We also assessed interaction terms, in particular, between the polymorphic genotypes and exposures to ABC. A traditional model for the probability of a patient having UADT cancer was built using backward elimination based on Akaike's information criterion.

To further assess if our epigenetic markers improved on or added value to the traditional methods, we added the epigenetic markers individually to the final traditional model and determined the independence and magnitude of the contribution of the epigenetic markers after adjusting for the effects of traditional risk factors. A combined model was constructed based on Akaike's information criterion. The significance of each predictor is presented as an OR and a 95% CI. A 2-tailed P < 0.05 indicated statistical significance. All statistical analyses were done using the SAS software system, version 9.2 (SAS Institute Inc.).

The performance of the above predictive models was compared by graphically displaying their relative performances on a receiver operating characteristic (ROC) curve that plotted the sensitivity against the false-positive rate (i.e., 1-specificity) over a range of cutoff values. The area under the curve (AUC) measured the performance of the test in correctly distinguishing subjects with and without cancer and the corresponding 95% CIs were calculated using 1,000 bootstrap sample sets (drawn with replacement) of the same size of the original dataset (31). The model calibration was assessed using the Hosmer–Lemeshow test for goodness of fit, which evaluated whether the predicted probabilities agreed with the observed probabilities in deciles. The bootstrap method evaluated statistical differences between the AUCs of epigenetic, traditional, and combined models over the same training set. We determined the optimal cutoff value for each model by selecting the point closest to theoretically perfect performance in discrimination and calculated the corresponding sensitivity and specificity pair.

Finally, to validate the epigenetic model, we recruited an independent validation set of a sample size similar to the training set at the National Taiwan University Hospital and quantified their DNA methylation levels in normal esophageal mucosae in a similar manner. We fitted an epigenetic model using the training set, saved the details of the fitted model such as the parameter estimates and SEs, and tested its performance on the validation set. The derived AUC was compared with that based on the training set using the z test.

The sample size of this prospective study was planned previously based on our previous study (30). That study found that methylation levels in the normal esophageal mucosae of cancer patients (i.e., group 3) were 2.7% ± 2.5% (average ± SD; median: 2.4%; range: 0%–16.3%) for HOXA9, 7.2% ± 5.1% (6.2%; 0.4%–20.3%) for NEFH, 1.8% ± 3.6% (0.9%; 0%–22.4%) for UCHL1, and 1.9% ± 1.7% (1.3%; 0%–6.8%) for MT1M. However, the methylation levels of healthy controls (i.e., groups 1 and 2) were not available. To test the hypothesis that methylation levels in the normal esophageal mucosae could differentiate cancer patients from noncancer subjects, we used the initial 10 healthy controls in the pilot experiment to estimate the required sample size; their average methylation levels were 1.2% for HOXA9, 0.2% for NEFH, 0.2% for UCHL1, and 0.6% for MT1M. To detect a significant difference between the case (i.e., group 3) and control (i.e., groups 1 and 2) at α = 0.05, β = 0.2, and a single-tail hypothesis, it was estimated that a total of 64 samples for each group would suffice using the Student t test. To compare the model based on epigenetic markers with the one based on traditional risk factors, we set a minimally acceptable level of accuracy at 80%: that is, a level that is commonly considered a prerequisite for population-based screening. Knowing that the average accuracy of traditional methods was approximately 70% (5, 20), an overall sample size of 231 was required to detect this difference at α = 0.05 and β = 0.2 using the χ² test. The sample size estimation was done using the nQuery Advisor 7.0.

Between January 2008 and April 2009, 272 consecutive patients were screened. Seventeen (6.3%) subjects were excluded, including 8 having a history of bleeding tendency, 2 with failed endoscopic examination, and 7 refusing endoscopic biopsy. The remaining 255 patients were recruited to the study as the training set. The demographic and clinicopathologic characteristics between cases and controls are shown in the Table 1. Histories of ABC were confirmed in 190 (74.5%) subjects; among them, 66 (34.7%) were double positives and 84 (42.2%) were triple positives. Of the 146 (57.3%) patients diagnosed with SCC of the UADT, 49 were diagnosed with head and neck cancer (including 4 with oral cavity cancer, 2 with laryngeal cancer, and 43 with hypopharyngeal cancer), 91 with esophageal cancer, and 6 patients with synchronous head and neck and esophageal cancers. Of the 109 subjects who had undergone the similar diagnostic procedures but had negative results, 46 (42.2%) had no history of ABC while 63 (57.8%) had used at least one of these 3 risk factors.

A total of 397 biopsied specimens were obtained in the training cohort and DNA methylation levels were successfully quantified in 372 (93.7%) of them. The methylation levels of all 4 epigenetic markers increased in a stepwise manner (P < 0.001) based on the exposure to alcohol, betel quid, or cigarettes (Fig. 1). The normal esophageal mucosae of healthy subjects not exposed to alcohol, betel quid, or cigarette (group 1) had methylation levels of 2.8% ± 3.2% (average ± SD; median: 1.8%; range: 0.3%–15.4%) for HOXA9, 0.4% ± 0.7% (0.2%; 0%–4.1%) for NEFH, 0.3% ± 0.5% (0.2%; 0%–2.4%) for UCHL1, and 0.9% ± 1.7% (0.4%; 0.1%–10.7%) for MT1M. The normal esophageal mucosae of healthy subjects with at least one of the above exposures (group 2) had methylation levels of 4.8% ± 5.7% (2.7%; 0.2%–29.3%) for HOXA9, 0.6% ± 1.1% (0.3%; 0%–7.4%) for NEFH, 1.2% ± 3.3% (0.3%; 0%–23.7%) for UCHL1, and 1.0% ± 0.7% (0.9%; 0%–3.1%) for MT1M. The normal esophageal mucosae of cancer patients (group 3) had methylation levels of 9.2% ± 9.9% (5.8%; 0%–65.8%) for HOXA9, 0.7% ± 1.0% (0.4%; 0%–7.3%) for NEFH, 2.0% ± 5.0% (0.4%; 0%–42.8%) for UCHL1, and 1.4% ± 2.2% (0.9%; 0.1%–18.0%) for MT1M.

Methylation levels in the normal esophageal mucosae correlated to cumulative lifetime exposures to ABC (Supplementary Figs. S1–7). Notably, a dose-dependent relationship was noted between the methylation levels of HOXA9 and alcohol consumption (P < 0.01). The methylation levels of UCHL1 and MT1M were higher in drinkers (P < 0.05). Although we did not identify a linear relationship between the methylation levels and dosage of betel quid use or cigarette smoking, betel quid users had significantly higher HOXA9 methylation levels than nonusers (P = 0.01) and smokers had consistently higher methylation levels for the 4 markers (P < 0.05). Regarding the concurrent use of ABC, the methylation levels of 4 marker genes were higher in individuals who had at least one risk factor (P ≤ 0.01). The methylation levels of HOXA9 increased in a stepwise manner with the number of exposures (P < 0.01), suggesting an additive effect.

When the data was stratified by the presence of numerous LVLs, which are a well-recognized indicator of SCC risk, higher methylation levels were noted in the cases with numerous LVLs (Supplementary Fig. S8), and the levels of the HOXA9 and UCHL1 genes reached statistical significance (P < 0.01). Stratifying the data by infection with H. pylori or human papillomavirus did not yield significant differences in the methylation levels of noncancerous mucosae.

The cancerous mucosae (group 4) had methylation levels of 33.2% ± 24.7% (28.8%; 0%–87.6%) for HOXA9, 6.4% ± 10.0% (2.3%; 0%–54.9%) for NEFH, 18.1% ± 23.7% (3.9%; 0%–91.4%) for UCHL1, and 11.5% ± 16.7% (2.0%; 0.1%–73.9%) for MT1M, which was a broad distribution with higher average values than each of the 3 normal groups (P < 0.001). Cancers of the head and neck and of the esophagus had similar methylation levels (Supplementary Fig. S9). No association was found between methylation levels and clinicopathologic characteristics, including TNM staging, tumor differentiation, and multiplicity (data not shown).

Methylation levels in the normal esophageal mucosae were significantly associated with the risk of SCC for the 4 epigenetic markers (Table 2). To predict head and neck cancer, AUCs for each of the epigenetic markers were 83% for HOXA9 (bootstrap 95% CI: 77%–88%), 69% for NEFH (60%–74%), 67% for UCHL1 (59%–74%), and 74% for MT1M (67%–81%). To predict esophageal cancer, similarly, the AUCs were 78% for HOXA9 (73%–83%), 71% for NEFH (65%–77%), 73% for UCHL1 (67%–78%), and 71% for MT1M (65%–77%). Overall, to predict UADT cancer, the AUCs for each of the markers were 80% for HOXA9 (75%–84%), 69% for NEFH (64%–75%), 71% for UCHL1 (66%–76%), and 72% for MT1M (66%–77%). Notably, HOXA9 showed the best discriminative ability and an additional dichotomous analysis showed that OR was 10.189 (5.430–19.118) when the cutoff point of the HOXA9 methylation level was set at 9.2% (the average value of cancer group).

An optimal epigenetic model to predict UADT cancer was constructed with 2 predictors of HOXA9 and NEFH methylation levels (AUC: 83%; bootstrap 95% CI: 79%–86%). The model had similar performances for head and neck cancer and for esophageal cancer (Supplementary Fig. S10). The sensitivity, specificity, and AUC were 76% (65%–87%), 75% (67%–83%), and 85% (80%–90%), respectively, in predicting head and neck cancer, and 72% (63%–81%), 73% (65%–81%), and 82% (77%–87%), respectively, in predicting esophageal cancer.

Thus, we confirmed that our 4 epigenetic markers were robust as a quantitative measure for field cancerization along the UADT.

The traditional risk factors were compared between cases and controls using univariate logistic-regression analyses (Table 2). Both groups were similar in age. The patients with SCC tended to be male, lower in BMI, and more likely to have the ABC habits. Serologically, their MCVs were higher, but their rate of H. pylori and human papillomavirus infections was similar to that of the controls. Endoscopy found that the cancer patients were more likely to have numerous LVLs in the background mucosae. A significant interaction was found between carriers of the genetic polymorphisms of an inactive ALDH2*2 allele and levels of alcohol consumption (OR: 1.212, 95% CI: 1.092–1.344, P < 0.001), as observed in the Hardy–Weinberg equilibrium, indicating that this genotype modified alcohol-related cancer risk.

Model performance was compared in terms of ROC curves (Supplementary Fig. S11). Overall exposure to ABC had a higher sensitivity (93%, 95% CI: 90%–96%) but a lower specificity (45%, 95% CI: 35%–54%), while the presence of endoscopic LVLs had a lower sensitivity (50%, 95% CI: 44%–56%) but a higher specificity (92%, 95% CI: 88%–98%). The AUCs increased as the model was based on MCV; then that based on overall exposure to ABC, endoscopic LVLs, and MCV (the traditional model; goodness-of-fit test, P = 0.77); then that based on methylation levels of HOXA9 and NEFH (the epigenetic model; goodness-of-fit test, P = 0.18); and then that based on the overall exposure to ABC, endoscopic LVLs, and methylation levels of HOXA9 and NEFH (the combined model shown in Table 3; goodness-of-fit test, P = 0.70), which was the most accurate. The sensitivity and specificity pairs of the optimal cutpoints were 56% (95% CI: 48%–64%) and 57% (48%–66%) for the MCV-based model, 74% (67%–81%) and 74% (66%–82%) for the traditional model, 74% (69%–79%) and 75% (67%–83%) for the epigenetic model, and 82% (76%–88%) and 81% (74%–88%) for the combined model. The performance of epigenetic model (AUC: 83%, bootstrap 95% CI: 79%–87%) was similar to that of the traditional model (AUC: 80%, 95% CI: 67%–91%; P = 0.51 for the comparison). After adding the epigenetic markers, the combined model (AUC: 91%, 95% CI: 88%–94%) was more accurate than the traditional model (P < 0.001).

Between May 2009 and May 2011, a total of 224 patients who underwent endoscopic screening were recruited as the validation set. They included 74 with negative endoscopies, 86 with head and neck cancer (6 with oral cavity cancer, 6 with oropharyngeal cancer, 5 with laryngeal cancer, and 69 with hypopharyngeal cancer), 59 with esophageal cancer, and 5 patients with synchronous head and neck and esophageal cancers. Their distribution of age (58.2 ± 12.5 y), gender (86.6% men), BMI (22.2 ± 3.7 kg/m²), and exposure to ABC (80.4%) was similar to that of the original training cohort. Methylation levels of the 4 markers in the normal esophageal mucosae of the control group were 2.9% ± 2.9% (average ± SD; median: 1.8%; range: 0%–12.4%) for HOXA9, 0.4% ± 0.7% (0.1%; 0%–6%) for NEFH, 0.5% ± 0.7% (0.1%; 0%–3.2%) for UCHL1, and 0.2% ± 0.5% (0.1%; 0%–3.4%) for MT1M. The normal esophageal mucosae of the cancer group had methylation levels of 13.5% ± 16.8% (5.9%; 0.1%–84.6%) for HOXA9, 1.1% ± 4.2% (0.2%; 0%–47.6%) for NEFH, 3.8% ± 11.7% (0.6%; 0%–82.0%) for UCHL1, and 2.4% ± 10.2% (0.2%; 0%–92.3%) for MT1M. When the epigenetic model fit in the training set was applied to the validation set, the AUC was 80% (bootstrap 95% CI: 73%–85%; goodness-of-fit test, P = 0.16); the difference in discrimination between the training and validation sets was not statistically significant (P = 0.43; Fig. 2). The HOXA9 remained the best discriminative marker with 78% AUC (95% CI: 72%–84%) in predicting UADT cancer.

Our present study quantified use of epigenetics for cancerization with small diagnostic biopsies from normal esophageal mucosae. We confirmed that this approach is equivalent to the traditional risk factor method, which typically requires a comprehensive set of clinical data. The biological significance of the epigenetic approach is supported by studies investigating the pathogenesis of SCC, which find that HOXA9 protein expression is deregulated in SCC (32), that oncogenesis is triggered by silencing NEFH to activate the Akt/β-catenin pathway (33), that the prognosis of SCC is worsened by hypermethylating UCHL1 (34) that interacts with p53 through the deubiquitination pathway (35), and that overexpressing MT is associated with metastatic and proliferative activity (36). None of these 4 genes had previously been correlated with carcinogen exposure or with cancer risk prediction.

Our study provided abundant information regarding the effects of traditional risk factors. As expected, ABCs were strong determinants that led to a lower BMI and a higher MCV, indicating poor nutritional status (7). However, judgment-based merely on the overall exposure to ABC was insufficient because the false-positive rate (55%) was high. Both of the infectious pathogens H. pylori and human papillomavirus were not associated with SCC risk (37, 38). Although several risk alleles have been previously identified (10–13), they served as effect modifiers rather than direct causes. Although the value of endoscopic LVLs has been confirmed, the usefulness of Lugol staining as a first-line screening tool is limited by a lower sensitivity (50%), in addition to significant side effects. A substantial proportion of subjects with cancer risk would be lost from surveillance if the risk assessment was solely based on the presence of endoscopic LVLs. However, the above limitations of traditional risk factors might support the credibility of our findings since our findings are similar to other literature reports (summarized in the Supplementary Table S2; refs. 15, 19, 20).

Taking into consideration both measured and unmeasured risk factors/genetic susceptibilities, molecular information derived from histologically normal mucosae has been extensively investigated, such as the field effects of oral leukoplakia, Barrett's esophagus (39), and colon neoplasm (40, 41). Mutations of p53 or p53 oncoprotein are relatively late events in carcinogenesis (42, 43), so, despite their biological plausibility as cancer predictors, they are insufficient for risk assessments for SCC in the UADT. Because aberrant DNA methylation usually precedes neoplastic transformation, qualitative studies have suggested that the presence of methylated genes may increase in a sequence from normal mucosa to precursor lesions to SCC (44–46). Our study further found that this phenomenon is quantitative for individual gene markers.

Research on clinical applications found that a candidate gene approach measuring the methylation of AHRR, p16INK4a, MT1G, and CLDN3 had unsatisfactory sensitivity of 50% and specificity value of 68% when screening for SCC in China with esophageal-balloon cytology (47). In contrast, the performance of our 4-gene panel may effectively identify subjects who will benefit from endoscopic surveillance. Our panel's superiority may be due to our meticulous search for candidate genes by oligonucleotide microarray, which decreased bias in choosing targets (30), and to using qMSP to extract full information without truncating the methylation value (23). Imaging-enhanced endoscopies minimized the risk of contamination during sample collection (28, 29). The similar methylation patterns between the 2 ethnically different populations from Japan and Taiwan further strengthens the generalizability of our approach. Our approach is accurate, inexpensive with the cost estimated at one USD per specimen marker, and high throughput. This can allow surveillance with sophisticated technologies to focus on high-risk patients, which is potentially cost effective.

Intriguingly, the HOXA9 gene, which is essential for embryogenesis and organogenesis, was identified as the most informative marker. This finding underscored the significance of HOX genes in regulating cell proliferation and differentiation, which is a balance that shifts in tumorigenesis (48). Also noteworthy are the contributions of overall exposure to ABC and of endoscopic LVLs, which remained significant in the combined model that yielded the highest AUC of 91%. This result may partially reflect the design of our screening procedure on the basis of alcohol and cigarette use (30) rather than on betel quid chewing. Betel quid chewing is a well-recognized class-1 carcinogen, so residual confounding cannot be ruled out. Recruiting more patients with heavy exposure to betel quid would be worthwhile because markers tailored to this substance could be identified by methylated DNA immunoprecipitation microarray (49).

Our risk-assessment models predicted the risk of actually finding SCC at the endoscopy when the biopsies were taken. The results may need to be verified by longitudinal data following the development of aberrant methylation in those with ongoing exposure to carcinogens. This would allow us to elucidate how many individuals with positive methylation markers would eventually go on to develop SCC (50). Individuals with normal mucosae, low methylation levels, and exposure to carcinogens may also benefit from risk stratification by modifying lifestyle risk factors to reduce cancer risk. In addition, although similar etiologic factors are involved in the development of head and neck cancer and esophageal cancer, they are not created equally. Our patients with head and neck cancer were mainly composed of those with hypopharyngeal cancer, so our results may not be so readily applied to those with oral cavity cancer. Therefore, this topic warrants further investigation.

In conclusion, we successfully quantified the field for cancerization using epigenetic markers. Our work provides a better model for risk assessment and may potentially be generalized across the high-risk regions to allocate the limited endoscopic resources for the surveillance and early detection of UADT cancers.

No potential conflicts of interest were disclosed.

The authors thank Luan-Yin Chang, MD, PhD, Department of Pediatrics, College of Medicine, National Taiwan University, for helping to determine the cutoff point to positively determine human papillomavirus infection.

This study was supported by research grants from the National Science Council (NSC96-2314-B-002-092-MY3) and the Taipei Institute of Pathology.

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.