Forecast of peak attainment and imminent decline after 2017 of oral cancer incidence in men in Taiwan

Oral cancer is the fourth most common cancer among men in Taiwan. The age-standardized incidence rate of oral cancer among men in Taiwan has increased since 1980 and became six times greater in 2014. To enable effective public health planning for oral cancer, research on the projection of oral cancer burden is essential. We conducted an age-period-cohort analysis on the incidence of oral cancer among men in Taiwan from 1997 to 2017 and extrapolated the trend to 2025. We found that the period trends for young adults aged between 25 and 44 have already peaked before 2017; the younger, the earlier, and then the trends declined. The cohort trends have peaked roughly at the 1972 birth cohort and then declined for all ages. Despite the increasing trend in the age-standardized incidence rate for oral cancer among men in Taiwan from 1997 to 2017, we forecast a peak attained, an imminent decline after 2017, and a decrease of 8.4% in age-standardized incidence rate from 2017 to 2025. The findings of this study contribute to developing efficient and comprehensive strategies for oral cancer prevention and control.

Age-period-cohort model. The APC model was used to analyze the oral cancer incidence rate and project the trend to 2025, given the strong cohort effect in Taiwan 36 . However, because of the perfect linearity of the temporal variables (cohort + age = period), an infinite set of parameter estimates with equal goodness of fit exists, causing the non-identifiability problem. In addition, we used data provided in 5-year age groups and 1-year periods in the study, which might cause additional identifiability issues with the unequal intervals in their definition of cohort indices 41 . To circumvent the non-identifiability problem inherent in the APC model, the linear cohort effect was not estimated in the APC model, presented as follows: where g(.) was the link function, µ was the expected incidence cases, m was the person-years, a, p, and c were the age variable, the period variable, and the cohort variable, respectively, θ was the intercept, α 1 and β 1 were the linear age effect and the linear period effect, respectively, and f A (a) , f P p , and f C (c) were functions of the age variable, the period variable, and the cohort variable, respectively, denoting the nonlinear effects. Maximum likelihood estimation was used for estimating the parameters in the APC model.

Ensemble learning and model selection.
We applied an ensemble technique to obtain an APC model with the best predictive performance. In this technique, various APC models in the ensemble are trained on the given training dataset. Finally, the model that performs best on the validation set is chosen for future use. We considered a threefold validation, splitting the complete data (21 calendar years from 1997 to 2017) into a training set (14 calendar years from 1997 to 2010) and a validation set (7 calendar years from 2011 to 2017). A total of 52 types of APC models (with the formula presented in Table S4) were considered for training. The model types referred to in previous research [11][12][13][42][43][44][45] , included polynomial APC prediction models (Type 1 to Type 17) 42 , Tzeng and Lee's APC prediction model (Type 18 to Type 22) 43,44 , and cubic splines APC prediction models (Type 23 to Type 52) [11][12][13]45 . For the cubic splines APC models, the knot locations were placed at: where n a , n p , and n c were the number of age groups, period groups, and cohort groups, respectively, and k a , k p , and k c were the number of knots for age, period, and cohort, respectively. We considered five types of link functions (log, power 2, power 3, power 4, and power 5) for the 52 model types. Also, with the assumption that historical trends will not continue indefinitely [11][12][13]42 , each APC model projection was applied to 21 levels (0%, 5%, 10%, 15%, …, or 100%) of year-on-year attenuation ( P * = (O mark − P) × A + P) , where P * was the attenuated for j = 1, 2, . . . , k p k c,k = n c k c +1 × k, for k = 1, 2, . . . , k c |Pp,a−Op,a| (Pp,a+Op,a)/2 , where P p,a was the predicted value and O p,a was the observed value) and the logarithmic score ( LOGS = 2017 p=2011 12 a=1 log Pr O p,a , where Pr (.) was the probability density function and O p,a was the observed value). We selected a final model with the lowest SMAPE or the maximal LOGS. Finally, we re-estimated the model parameters based on oral cancer incidence data from 1997 to 2017 (all available data) and made projections for 2025. We presented a diagram of the method for building APC prediction models in Fig. S1.
We also conducted sensitivity analyses and presented results in the supplementary materials. The sensitivity analyses included working the model selection with a fivefold validation (16 calendar years from 1997 to 2012 for training and the remaining five calendar years for validating), applying log-transformation on the SMAPE where t ≥ 2012 , and performing the same analysis for women. All statistical analyses were performed using SAS version 9.4 (SAS Institute Inc, Cary, NC, USA). The SAS code for analysis was presented in the supplementary materials.

Results
We selected the optimal model as a cubic splines APC model (type 26 in Table S4) with 80% attenuation, presented as follows: log  Fig. 1. The age-standardized incidence rate increased considerably from 1997, with the incidence rate doubled in 2009 compared with 1997, but then the increasing trend slowed. The projection indicated that the trend would start to decline in 2017 and further decline until 2025. Data on age-specific oral cancer incidence rate among men in Taiwan by calendar year and birth cohort are displayed in Fig. 2. Period trends diverged across age groups. Period trends for older age groups (70-74, 75-79, and 80-84) were rapidly increasing, and the projections also show an increasing trend. Period trends for the middle-age groups (40-44, 45-49, 50-54, 55-59, 60-64, and 65-69) were increasing at first but then leveled off, with projections indicating a stable trend (60-64 and 65-69) or decrease (40-44, 45-49, 50-54, and 55-59). Period trends for the younger age groups (25-29, 30-34, and 35-39) initially increased rapidly, peaked-the younger, the earlier, and then decreased. Period trends for these younger age groups are projected to further decline until 2025. Birth cohort trends were consistent for all age groups: trends increased before 1973 and decreased after that.
The age-standardized and age-specific oral cancer incidence per 100,000 population (observed rates in 2017 and projected rates in 2025) and projected percentage changes from 2018 to 2025 are shown in Table 1. The agestandardized incidence rate is poised to decrease by 8.4% from 2017 to 2025. The age-specific incidence rates of age groups 80-84, 75-79, and 70-74 are predicted to increase substantially by ≥ 20% from 2017 to 2025. The age-specific incidence rates for 65-69 and 60-64 are expected to increase slightly by < 10% from 2017 to 2025. The age-specific incidence rate of 55-59 will decrease by approximately 10% from 2017 to 2025. The age-specific incidence rates of 50-54 and 35-39 will reduce by about 15% from 2017 to 2025. The age-specific incidence rates of the remaining age groups (45-49, 40-44, 30-34, and 25-29)   www.nature.com/scientificreports/ age-standardized incidence trend to men. However, the age-specific incidence rate among women increased slower than among men for 65-84.

Discussion
The age-standardized incidence rate of oral cancer among men in Taiwan presented a rapidly increasing trend at first, but the increasing trend slowed after 2009. The historical trend of age-standardized rate alone makes it challenging to visualize peak attainment and imminent decline after 2017. By contrast, the APC analysis showed that period trends for young adults aged 25-44 years had already peaked before 2017-the younger, the earlier, and then the trends declined. Moreover, a significant cohort effect was revealed, decreasing incidence rates at roughly the 1972 birth cohort for all ages. The combined results indicate a peak has been attained, and the trend is now for oral cancer incidence to decrease among men in Taiwan. Previously, Hsu et al. 30 predicted a future decline but not before 2025. However, they only created one APC model and did not use techniques to prevent overfitting. By contrast, we used a data-splitting method and constructed an ensemble APC model to allow for flexible model specification. Furthermore, Hsu et al. 30 only studied the trend in age-standardized incidence rate but not the period trends and birth cohort trends by age. We did not estimate the linear cohort effect for making parameters identifiable in the study. The use of additional constraints can successfully decompose the APC effect. However, different constraints may cause drastically different or even contradictory results. There is no consensus in the APC literature on which constraints are the best and used. Besides, the incidence projections in this study were impervious to the non-identifiability problem because the fitted values in the nonidentifiable APC model were the same for all possible sets of parameter estimates. According to Figs. 1 and 2, a discontinuity could be observed when the prediction started, caused by imposing the 80% attenuation on the predictive value. The idea of attenuating the projected trend comes from the belief that historical trends will not continue indefinitely and was shown to be valuable empirically for making future predictions by Møller et al. [11][12][13]42 When alleviating the attenuation level from 80 to 0%, the result showed a rapidly decreasing trend after 2017 (Fig. S8). Still, the SMAPE would increase from 8.59 to 11.52, and the LOGS would drop from − 469.75 to − 490.82. With the experience by Møller et al. 13,42 , we also found that it worked better to gradually reduce the projected trend (SMAPE = 8.59 and LOGS = − 469.75) than attenuate the trend a period after the last observation (SMAPE = 9.04 and LOGS = − 486.84). Therefore, it seems favorable to gradually attenuate the projected trend in the study. In addition, the selected type 26 model with 80% attenuation had the lowest SMAPE (8.59) but the second-highest LOGS (− 469.75). Notably, the same model with a bit higher 85% attenuation had the second-lowest SMAPE (8.78) and the highest LOGS (− 460.61). Therefore, it is suitable to predict the oral cancer incidence among men in Taiwan with the type 26 model with 80% or 85% attenuation.
The data splitting method in the study was referred to in the previous studies 20, 25 in Taiwan. However, unlike a 2-fold validation in those studies, we applied a 3-fold validation, two-thirds of the calendar years (1997-2010) used for training and the remaining (2011-2017) for validating, for the following motivation and reasons: (1) given the calendar year ended in 2010, we could include enough data cells for the 1972 birth cohort and later for training (Table S3) to capture the cohort effect for oral cancer among men in Taiwan 36 , (2) a sensitivity analysis with a 5-fold validation was considered, and the similar projected trends were obtained (Figs. S6-S7), and (3) Taiwan have implemented the National Health Insurance Program and passed the legislation of Tobacco Hazards Prevention Act since 1997, which might impact the oral cancer incidence trend in Taiwan.
Betel quid chewing and smoking are risk factors for oral cancer among men in Taiwan 46,47 . In 2005 in Taiwan, the betel quid chewing rate was approximately 10% 48 , and the smoking rate among men was about 50% 49 . Table 1. Age-standardized and age-specific oral cancer incidence per 100,000 population (observed rates in 2017 and projected rates for 2025) and projected percentage changes from 2018 to 2025. # The World Health Organization 2000 World Standard Population was used to compute the truncated age-standardized incidence rates (age 25-84 years), see Table S6. www.nature.com/scientificreports/ Notably, the betel quid chewing rate for the indigenous peoples of Taiwan was four times the overall rate in Taiwan 50 . The World Health Organization Global Oral Health Programme presented the common risk factor approach for preventing and controlling non-communicable diseases 51,52 . Public health interventions, including tax policies and health education 49,53-56 , for preventing excessive risk exposure of these lifestyle factors 57 have been implemented since 1997 in Taiwan. These measures have continuously declined Taiwan's betel quid chewing and smoking rates 49,54 . Su et al. 36 conducted an APC analysis to examine the incidence rate of oral cancer among men in Taiwan from 1997 to 2016. They found a strong association between the cohort effect on oral cancer incidence rate among men and average betel nut consumption with a lag time of 30 years. A decrease in smoking rates may also contribute to the decreasing oral cancer incidence rate after the 1972 birth cohort. Per our projections, by 2025, the age-specific incidence rate will have decreased for those between 25 and 59 years old but increased for those between 60 and 84 years old. Health care professionals have actively implemented long-term care services since 2007 to alleviate the excessive burden of an aging population in Taiwan to ensure older adults can live healthier lives 58,59 . Early detection of oral cancer can prolong life expectancy 60 . Taiwan started a national oral cancer screening program in 2004 61,62 that provides a free oral mucosal examination every two years for Taiwanese residents with habits of smoking or betel quid chewing. Between 2004 and 2009, the overall oral cancer screening rate was 55.1% in Taiwan 62 . A shift has been observed, indicating that patients with malignant oral cancer can be detected earlier because of the screening program 62 . In addition, we obtained contrasting projected oral incidence trends in men and women for 65-84. Because men might have a higher prevalence of risk factors and more opportunities to be screened for oral cancer, the early detection due to the screening program might have caused the increase in oral cancer incidence among men aged between 65 and 84 since 2004.
In conclusion, despite the increasing trend in the age-standardized incidence rate of oral cancer among men in Taiwan from 1997 to 2017, we determined that a peak was reached in 2017, and the incidence has subsequently been in decline, with a decrease of 8.4% in age-standardized incidence from 2017 to 2025. Therefore, the findings of this study contribute to developing efficient and comprehensive strategies for oral cancer prevention and control.

Data availability
The authors confirm that all data underlying the findings are fully available without restriction. The data underlying the results of this study are either available in the manuscript or upon request from the corresponding author, Wen-Chung Lee. The raw data cannot be made available in supplemental files or a public repository because of privacy or ethical reasons. However, the dataset used for analysis can be obtained in an anonymized form upon request from the corresponding author.