Recalibrating Risk Prediction Models by Synthesizing Data Sources: Adapting the Lung Cancer PLCO Model for Taiwan

This article is featured in Highlights of This Issue, p. 2093

Absolute risk models estimate disease risk in an upcoming time interval based on known risk factors for a healthy individual in a population, accounting for competing causes of death (1, 2). Absolute risk models have important clinical and public health applications (3, 4). Strategies to develop, validate, and update absolute risk models have been important research topics in recent decades (2, 5, 6). Although prospective cohorts are ideal for their development, validation, and updating, the required sample size would be large and follow-up periods long if the disease incidence rate is low, such as cancers at specific sites. Methods that synthesize multiple data sources without using prospective datasets have been proposed for model development following the seminal contribution of Gail and colleagues (1, 7–11).

Indeed, both Gail and colleagues and Costantino and colleagues combined estimates of relative risks associated with certain risk factors and estimates of the baseline hazard and attributable risk to obtain estimates of the probability of developing breast cancer, using competing risk models. The modification by Costantino and colleagues used age-specific invasive breast cancer rates and attributable risk estimates from the Surveillance, Epidemiology, and End Results rather than from the Breast Cancer Detection Demonstration Project (1, 8). Recognizing a growing demand to develop and apply models for absolute risk prediction, the iCARE package builds competing risk models by synthesizing multiple data sources containing information on relative risks, the distribution of risk factors in the population, and age-specific incidence rates (11).

Chien and colleagues developed logistic regression models for predicting lung cancer occurrence in the upcoming 6 years among never-smoking Taiwanese females, based on an age-matched case–control study (AMCCS) and the age-specific 6-year lung cancer incidence rates (ASSIR) for never-smoking females in Taiwan (10). The AMCCS was used to estimate the effects of risk factors other than age and the intercept; given these effect estimates, they used the ASSIR and risk factor distributions among the controls to estimate the age effect and intercept. The AMCCS was obtained from a case–control study of lung cancer; ASSIR, accounting for competing causes of death, was estimated using the Taiwan Cancer Registry (TCR), the Taiwan Cause of Death Database (TCOD), age-specific population size, never-smoking rates in the female population and in female patients with lung cancer, and the Taiwan life table.

Because validating and updating risk models are essential toward better risk prediction models (12–14) and except for discrimination, are currently carried out using prospective cohorts, we aimed to propose methods for validating or adapting absolute risk models for another population by synthesizing multiple data sources, when no suitable prospective cohorts are available. This would help alleviate the disparities due to risk models, with or without incorporating polygenic risk scores (15, 16).

To make the presentation concrete, we exemplified the methods by adapting the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial 2012 model (PLCO_M2012) for Taiwan. It is a logistic regression model that estimates the probability of a smoker developing lung cancer in a 6-year period, using age, ethnicity, education, body mass index (BMI), chronic obstructive pulmonary disease (COPD), family history of lung cancer, personal history of cancer, smoking status, average number of cigarettes smoked per day, years smoked, and quit time (17). It was used to guide the selection of participants for low-dose computed tomography (LDCT) lung cancer screening trials (18) and was prospectively validated in the United States, Germany, Australia, Canada, U.K., Brazil, and Poland (18–28). It was also used to assess the clinical utility of polygenic risk scores for risk stratification regarding LDCT screening (29). The National Comprehensive Cancer Network 2018 guidelines approve selection based on PLCO_M2012 risk (30). However, the validation or adaptation of the PLCO_M2012 model has not been reported in Asia.

To take advantage of the excellent performance of the PLCO_M2012 model in the west and to reduce the possibility of overfitting, we considered two parsimonious approaches to adapting: Approach 1 updated only the intercept to deal with the calibration-in-the-large problem and Approach 2 derived the overall calibration slope and then handled the calibration-in-the-large problem (12, 13). For these, we formed population datasets and derived summary statistics or information for Taiwan and then used them to recalibrate and assess the risk model. We (i) constructed an age-matched case–control study of ever-smokers (AMCCSE), (ii) estimated the number of ever-smoking lung cancer patients diagnosed in 2011–2016 (NESLP2011), and (iii) synthesized a dataset resembling the population of cancer-free ever-smokers in Taiwan at the end of 2010 (SPES2010) with respect to the risk factors in the PLCO_M2012 model. In this paper, a person is cancer-free at a time point in her/his life if she/he has never been diagnosed with any cancer before that time point, and a person is a cancer survivor at a time point in her/his life if she/he has been diagnosed with certain cancer before that time point.

For Approach 1, we changed the intercept by requiring the resulting total risk of individuals in the SPES2010 equal NESLP2011. For Approach 2, we decomposed its linear predictor into two parts: the intercept and age term and the remaining term. We first used AMCCSE to conduct the “calibration slope” step that recalibrated the remaining term. Given the calibration slope, we then performed the calibration-in-the-large step that recalibrated the intercept and the age term by requiring the resulting total risk of individuals in the SPES2010 equal NESLP2011. In either approach, we assessed the performance of the adapted model in terms of discrimination, subgroup calibration, and clinical usefulness (12) using SPES2010 and other datasets.

The ever-smoking patients with lung cancer in the AMCCSE were collected from the case–control component of the Taiwan Genetic Epidemiology Study of Lung Adenocarcinoma (GELAC) and the Taiwan Lung Cancer Pharmacogenomics Study (LCPG). The ever-smoking healthy controls were from the Taiwan Biobank (RRID: SCR_010557) and the case–control component of the GELAC. Limiting to the age range 50 to 74, we formed a total of 798 age-matched groups, where each group had exactly one case and one to five age-matched healthy controls, involving a total of 3,508 controls. Figure 1A presents the procedure formatting the AMCCSE, including the inclusion and exclusion criteria applied to the Taiwan Biobank. Supplementary Materials and Methods Texts S1–S4 have the details. Inclusion and exclusion criteria for individuals from the GELAC and LCPG were described in earlier publications (10, 31–33). Although blinding was not applied to this study, individuals in the Taiwan Biobank were deidentified before being provided to us.

We used the Taiwan Biobank, consisting of cancer-free individuals at recruitment, to construct the dataset SPES2010, which consequently included only cancer-free individuals. We first determined the age- and sex-specific numbers of cancer-free ever-smokers in the SPES2010. It was constructed on the basis of (D1) the age- and sex-specific Taiwanese population size at the end of 2010 using Monthly Bulletin of Interior Statistics (MBIS) from the Taiwan Ministry of the Interior (34); (D2) estimates of age- and sex-specific numbers of cancer survivors (cancer prevalence) in Taiwan at the end of 2010 using the linkage of the TCR, TCOD, National Health Insurance Research Database (NHIRD); (D3) the age- and sex-specific smoking rate for the year 2010 using the Taiwan Adult Smoking Behavior Survey (ASBS); (D4) the age- and sex-specific number of ever-smokers in the Taiwan Biobank having information on all the risk factors in the PLCO_M2012 model. The procedures leading to the estimates in datasets D2 and D3 are given below in the section on data sources. These datasets are included in Supplementary Tables S1–S3. For each age and sex, Supplementary Table S1 reports the smoking rates (percentage of ever-smokers) in the population; Supplementary Table S2 reports the cancer prevalence at the end of 2010.

Because a person was either cancer-free or a cancer survivor, we used D1 and D2 (Supplementary Table S2) to obtain age- and sex-specific cancer-free population sizes by subtraction. Assuming the age- and sex-specific smoking rates in the cancer-free population approximated those in the general population (Supplementary Table S1), we report in Supplementary Table S3 the estimated age- and sex-specific numbers of cancer-free ever-smokers; Supplementary Table S3A for females, Supplementary Table S3B for males, and Supplementary Table S3C for females and males combined. They determined the age- and sex-specific population size of the SPES2010. Given an individual in SPES2010, we assigned to this individual the risk-factor profile of an ever-smoker randomly selected from the Taiwan Biobank having the same age and sex and without missing information on the risk factors in the PLCO_M2012 model. This suggests that the age- and sex-specific distribution of these risk factors in the SPES2010 resembled those in the Taiwan Biobank. Figure 1B outlines the above procedure.

According to the TCR and TCR Long Form (TCRLF), approximately 83% of the lung cancer patients diagnosed in 2011–2016 reported whether they were ever-smokers or never-smokers; see Supplementary Table S4A. On the basis of the age-, sex-, and calendar year–specific smoking rates among patients with lung cancer derived from the TCRLF, we estimated the age-specific numbers of ever-smoking lung cancer patients in the TCR for 2011–2016 (Supplementary Table S4B). These were used to estimate, among those aged 50 to 74 at the beginning of 2011, the number of ever-smoking lung cancer patients diagnosed in 2011–2016. Supplementary Table S4B was also used to estimate the age-specific 6-year lung cancer incidence rates among ever-smokers (ASSIRE; Supplementary Table S4C).

Here, we only explain Approach 2, because Approach 1 is similar to the second step of Approach 2. Because AMCCSE was suitable for modifying the effects for all the risk factors other than age and the intercept and because we preferred a parsimonious approach, we decomposed the linear predictor of the PLCO_M2012 model (17, 35) into two components: its weighted sum of the intercept and the age effect, −4.532506+0.0778868 (Age - 62), is called the intercept-age factor. The remaining part of the linear predictor is called the non–intercept-age factor, which is a weighted sum of the effects of the other risk factors.

We fitted a logistic regression model with the intercept-age factor and the non–intercept-age factor only. Treating the former as the matching variable, we first fitted the logistic regression model, using a conditional likelihood approach (36), to the AMCCSE to obtain the OR of the non–intercept-age factor. Given the OR of the non–intercept-age factor from the first step, the second step estimated the OR of the intercept-age factor by requiring the resulting total risk of individuals aged 50 to 74 in the SPES2010 to be equal to NESLP2011.

The adapted model is called PLCOT-1 if Approach 1 is used and PLCOT-2 if Approach 2. Note that in this adaptation, what we need from SPES2010 was the distribution of the risk factors in the population and the population size. Details are in Supplementary Materials and Methods Text 5.

All computations are carried out using R language. The conditional likelihood approach to logistic regression was implemented using the clogit function in R, which was also used to obtain the 95% confidence interval (CI).

We assessed discrimination for PLCOT-1 by computing the area under the receiver operating characteristic curve (AUC) using all the cases and controls from the AMCCSE, which was not used in the PLCOT-1 adaptation.

We assessed discrimination for PLCOT-2 by bootstrap. On the basis of the AMCCSE, we obtained bootstrapping optimism-corrected discrimination in terms of AUC. Here one bootstrap sample was a set of age-matched case–control groups sampled from the AMCCSE with replacement and having the same sample size as that of AMCCSE. A total of 1,000 bootstrap samples were used and the correction method is detailed in Section 5.3.4, Steyerberg (6).

Because the TCR and TCRLF during 2011–2016 are follow-up data of the Taiwanese population at the end of 2010, comparison of SPES2010 and the TCR and TCRLF provided opportunities for assessing the calibration and clinical usefulness of the PLCOT models in terms of subgroups defined by age and smoking experiences. We considered sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for risk-based criteria as well as those defined by age and smoking experiences.

Consider, for example, the condition that individuals smoked ≥ 20 pack-years, smoked within the past 15 years, and were aged 50 to 74 at the beginning of 2011. We estimated the number of lung cancer patients diagnosed in 2011–2016 satisfied this condition using the information in the TCR and TCRLF. Because this information in the TCRLF pertained to time at cancer diagnosis, correction was properly made; details are provided in Supplementary Materials and Methods Text S6. We also estimated the total PLCOT risk of individuals satisfying the condition in the SPES2010. The ratio of the former (observed) to the latter (predicted) provided the calibration assessment on this subgroup. The predicted number was also used to study clinical usefulness. For example, we report sensitivity to be the ratio of the predicted number to NESLP2011, assuming the number of cases developed during 2011–2016 among the SPES2010 equaled NESLP2011. We considered here the predicted, rather than the observed, because we wanted to compare the performance of simplified criteria with the corresponding PLCOT risk-based criteria.

The following simplified criteria appeared in the literature. The 2013 US Preventive Services Task Force (USPSTF) criterion is smoking ≥ 30 pack-years, smoking within the past 15 years, and aged 55 to 80 (USPSTF13; ref. 37). The 2021 USPSTF criterion is smoking ≥ 20 pack-years, smoking within the past 15 years, and aged 50 to 80 (USPSTF21; refs. 38, 39). Another criterion was studied in the Nederlands-Leuvens Longkanker Screenings Onderzoek (NELSON) trial (40). One met this criterion if one was aged 50 to 74, smoked at least 10 cigarettes per day for at least 30 years or 15 cigarettes per day for at least 25 years, and smoked within the past 10 years. In this study we considered USPSTF21, NELSON, 40–10 (smoked 40 or more pack-years and smoked in the past 10 years), and 30–15 (smoked ≥ 30 pack-years and smoked in the past 15 years).

The TCR, launched in 1979, is a population-based cancer registry collecting information on newly diagnosed cancers at all hospitals in Taiwan with 50 or more beds. Its quality has been improving and was recently reviewed (41, 42). The TCRLF has included smoking information on patients with cancer since 2011. This study considered only lung cancers that were the first invasive cancers in patients.

The TCOD includes cause-of-death information for individuals in Taiwan since 1971. It is maintained by the Department of Statistics, Taiwan Ministry of Health and Welfare; its quality has been previously described (43). The TCOD adopted the national identification card number (NICN) in 1985. During 1985–2018, there were 4,398,359 unique death records that included individuals’ NICN, sex, birth date, death date, and cause of death. The original TCOD contains 4,405,868 records for this period; thus, < 0.2% of the data were excluded during data cleaning.

Taiwan's NHIRD was based on the administrative database of the National Health Insurance Program, which started in 1995 and has a coverage higher than 99% of its population of 23+ million. The NHIRD has been shown to be a valuable research resource (44).

Taiwan Biobank, started in 2008, is an ongoing community-based cohort of Taiwanese participants aged 30 to 70 who are cancer-free at enrollment and have information from basic physical examinations, questionnaires, and blood samples taken at enrollment. Information was also collected during follow-up appointments. This study used all the Taiwan Biobank data provided to us by November 2020, including a total of 122,071 participants; among them, 27,209 had follow-up data. Among those with follow-up data, 417 participants had a cancer diagnosis. To include participants older than 70 years, we used follow-up data for those who were cancer-free at the follow-up appointment. Figure 1A includes the data-cleaning process; details are provided in the Supplementary Materials and Methods. Chien and colleagues contains additional information (10).

Starting in 2007, the Taiwan ASBS reports the sample size and the proportion of ever-smokers for each survey by year, age, and sex. Using these, we obtained the number of ever-smokers in each survey and then calculated the "locally averaged" age-specific smoking rate for each year and sex. For example, the smoking rate for males aged 50 in 2010 was the proportion of ever-smokers among the male samples aged 49 to 51 in 2009–2011. Supplementary Table S1 reports the age- and sex-specific smoking rates for 2010.

Using the linkage of the TCR for 1979–2016, TCOD for 1985–2018, NHIRD for 2000–2017, we estimated the number of cancer survivors at the end of 2010. A cancer survivor is one included in the TCR for 1979–2010, not in the TCOD for 1985–2010, and in the NHIRD 2000–2010. Supplementary Table S2A presents the age- and sex-specific numbers of cancer survivors. Supplementary Table S2B reports the number of cancer survivors whose diagnoses were in the years between 1979 and 1988. Supplementary Table S2B suggests that the underestimation of cancer prevalence due to diagnoses before 1979 is likely minimal.

Cases in the GELAC were Han Chinese aged 18 or older with incident lung cancer diagnosed during 2000–2015 in Taiwan. No limitations on sex, smoking status, histology, or stage were imposed. The controls in the GELAC study were recruited from the health examination centers. The GELAC has been used to study lung cancer in never-smoking females and ever-smokers (10, 32, 33, 45). Supplementary Materials and Methods provides more information.

The LCPG recruited from health records late-stage lung cancer patients for whom epidermal growth factor receptor mutation statuses were available during 2015–2017 (10). More information about the LCPG is provided in Supplementary Materials and Methods. The structured questionnaires were administered to the GELAC and LCPG participants.

This study was approved by the institutional review board of the National Health Research Institutes in Taiwan (RRID: SCR_000335) and conforms to the Declaration of Helsinki provisions. All the datasets used in this study were provided to us after deidentification except GELAC and LCPG. All study subjects in the GELAC and LCPG provided signed informed consent prior to the commencement of this study.

The linkage of TCR, TCOD, and NHIRD can be performed and used for research upon approval of the Data Science Center, MOHW, Taiwan. The Taiwan Biobank dataset can be used for research upon approval of the Taiwan Biobank (https://taiwanview.twbiobank.org.tw/index). ASBS can be freely downloaded from the Health Promotion Administration, MOHW, Taiwan. Age-, year- and sex-specific population sizes can be freely downloaded from MBIS, Ministry of Interior, Taiwan. The use of datasets for the GELAC and LCPG studies need the approval of the NHRI IRB.

Using the Taiwan Biobank, GELAC, and LCPG, we followed the procedures in Fig. 1A to form the AMCCSE. Using the MBIS, ASBS, TCR, TCOD, and NHIRD, we followed the procedures in Fig. 1B to form the SPES2010. Table 1 presents the characteristics of the AMCCSE and SPES2010, in view of the risk factors in the PLCO_M2012 model. Supplementary Table S5 presents the smoking-related characteristics of these data sources. Table 1 shows that for these risk factors, their distributions among the AMCCSE cases were different from those among the controls, confirming that these were indeed risk factors for lung cancer among the ever-smokers in Taiwan. The characteristics of the SPES2010 shows that there were approximately 1,562,798 cancer-free ever-smokers aged 50 to 74 in Taiwan, and among them, more than 94% were males. These cancer-free ever-smokers accounted for approximately 27% of the Taiwanese population aged 50 to 74, which was approximately 5,765,938, according to Supplementary Table S5C. A comparison of Table 1 with the Supplementary Table S6 in Chien and colleagues (10) suggests that COPD and family history of lung cancer were more prevalent among ever-smokers than those among never-smoking females.

It follows from Supplementary Table S4B that among those aged 50 to 74 at the beginning of 2011, the number of ever-smoking patients with lung cancer diagnosed in 2011–2016 (NESLP2011) was estimated to be 17,374. Combined with the age-specific cancer-free ever-smokers at the end of 2010 reported in Supplementary Table S3C, we report in Supplementary Table S4C the ASSIRE.

Supplementary Table S5A indicates that, according to the TCRLF dataset, approximately 45% of the patients with lung cancer in Taiwan were ever-smokers (46); that only 36% in the GELAC and LCPG studies were ever-smokers probably reflects their recruitment criteria. The difference in the age and sex distribution reported in Supplementary Table S6 and Supplementary Table S5A suggests that some selection bias exists in the Taiwan Biobank.

By requiring that NESLP2011, being 17,374, equals the estimated total risk of individuals in SPES2010, we obtained the adapted model PLCOT-1, whose beta coefficients were the same as those of PLCO_M2012 except for the intercept. Table 2 presents these coefficients.

By fitting a logistic regression model with the intercept-age factor and the non–intercept-age factor defined by PLCO_M2012 to the AMCCSE, we first obtained the OR 0.514 for the non–intercept-age risk factor; given this, we obtained the OR 0.859 for the intercept-age factor by requiring that NESLP2011, being 17,374, equals the estimated total risk of individuals in SPES2010. This resulted in the adapted PLCOT-2 model. Table 2 also includes the beta coefficients of PLCOT-2. The OR of the non–intercept-age factor had 95% CI (0.443–0.585). Details are in Supplementary Materials and Methods Text S5.

The AUC for PLCOT-1 was 0.7776, and the bootstrapping optimism-corrected AUC of PLCOT-2 was 0.7549, and its apparent AUC was 0.7553.

Table 3 evaluates the PLCOT models using SPES2010 and NESLP2011. It reports the calibration assessment for both PLCOT-1 and PLCOT-2 on subgroups defined by 40–10, 30–15, USPSTF21, and NELSON and shows that PLCOT-1 had excellent subgroup calibration and that PLCOT-2 had good calibration.

For each of the PLCOT models, Table 3 also compares the sensitivity, specificity, PPV, and NPV of the four simplified LDCT lung cancer screening criteria with those of the four PLCOT risk-based criteria selecting the same number of individuals for screening. They show that the corresponding PLCOT risk-based criteria performed better than their simplified criteria counterparts in terms of these measurements. Consider USPSTF21, for example. A total of 518,239 (33.1%) individuals among all the 1,562,798 individuals in the SPES2010 satisfied the USPSTF21 criteria. The PLCOT-2 risk threshold 0.0124 would select the same number of individuals from SPES2010. The latter had a sensitivity of 66.7% and a PPV of 2.2%, while the former had a sensitivity of 58.7% and a PPV of 2.0%. Thus, the PLCOT model risk-based criteria would have potentially identified approximately 8% more cancers than the USPSTF21 criteria if the same number of individuals had been selected for screening.

To help communicate sensitivity and PPV, we present in Supplementary Figures S1 and S2 the Lorenz curves for the PLCOT-based risk distribution on the SPES2010, which plot predicted total lung cancer incidence against the number of individuals at highest risk (6).

Figure 2 presents the densities of risk among subgroups of SPES2010 by smoking levels, Figs. 2A and B regarding pack-years smoked and Figs. 2C and D regarding quit time. While Fig. 2 shows that both PLCOT models vary with smoking experiences, they suggest that PLCOT-1 risks vary more, in line with the fact that the OR of the non–intercept-age factor is less than 1. Indeed, for PLCOT-2, the mean 6-year risk was 2.7% for those who smoked 40+ pack-years, 1.7% between 30 and 40 pack-years, 1.5% between 20 and 30 pack-years, 0.6% less than 20 pack-years and the corresponding mean risks for PLCOT-1 were 3.8%, 2.0%, 1.4%, and 0.3%, respectively. Similar results hold for quit time.

For the clinically relevant 6-year lung cancer risk threshold of 0.0151 (35), Fig. 2A and B show that the vast majority of those who smoked 40+ pack-years had risks higher than 0.0151 and very few of those who smoked less than 20 pack-years had risks above this threshold. Figure 2C and D suggest that the proportion of high-risk individuals decreased persistently with the number of cessation years.

Figure 3 presents studies about smoking cessation effects on lung cancer risk. Figure 3A and B compare the densities of the PLCOT risks of lung cancer among 60-year-old current smokers in the SPES2010 who had smoked more than 10 years with more than 20 cigarettes per day with those of the same people if they had quit smoking when they were 50 years of age. Figure 3C and D compare the age-specific means of PLCOT risk of lung cancer for each age from 50 to 74 among the 50-year-old current smokers in SPES2010 who smoked at least 20 cigarettes per day with those among the same people if they had quit smoking at age 50. Figure 3A and B suggest a great risk reduction in terms of the 0.0151 threshold. Indeed, nearly no one had risks less than the threshold for either model; however, if they had quit smoking at 50, then 64% (55%) of them had risk less than 0.0151 under PLCOT-1 (PLCOT-2). Figure 3C and D suggest that the reduction in lung cancer risk increased considerably with quitting time, with PLCOT-1 reporting a much larger reduction.

This paper presented methods for adapting risk prediction models without prospective cohorts. Having formed an AMCCSE, estimated the number of ever-smoking lung cancer patients in 2011–2016, and prepared the dataset resembling the population of cancer-free ever-smokers at the end of 2010, we exemplified the methods by adapting the PLCO_M2012 model for Taiwan use. The PLCOT-1 model had an AUC of 0.78 and excellent performance in terms of subgroup calibration and clinical usefulness. The PLCOT-2 model had a bootstrapping optimism-corrected AUC of 0.75 and quite good performance in terms of subgroup calibration and clinical usefulness. Using these models, we reported risk distributions according to smoking exposure levels and described the effects of smoking cessation on risk reduction. To the best of our knowledge, the PLCOT models represent the first attempts to recalibrate or validate the PLCO_M2012 model in Asia.

In line with the literature that risk model updating methods range from simply updating the intercept to reestimating for each risk factor (5, 6), we synthesized datasets so that we could conduct updating at roughly three levels of sophistication. At the basic level, Approach 1 used SPES2010 and NESLP2011 to modify only the intercept to handle the calibration-in-the-large problem, resulting in PLCOT-1. At the second level, Approach 2 used the additional dataset AMCCSE to consider both the overall calibration slope and calibration-in-the-large problem, resulting in PLCOT-2. At the next level, we could replace the calibration slope step in Approach 2 by reestimating the effect of each risk factor other than the intercept and age, using the same AMCCSE. Following the approach in Chien and colleagues (10), we could also replace the calibration-in-the-large step in Approach 2 by reestimating the age effect and intercept using the age-specific 6-year lung cancer incidence rates among cancer-free ever-smokers, reported in Supplementary Table S4C. However, extensive model revision requires larger datasets in general. Indeed, we considered several such extensive model revisions and found poor subgroup calibration.

Although PLCOT-1 performed a little better than PLCOT-2 in terms of discrimination, subgroup calibration, sensitivity, specificity, PPV, and NPV, we presented both models in this paper because we intended to exemplify the methodology more fully, and we think future studies using prospective cohorts would give more conclusive comparisons. Indeed, we expect to conduct a validation study when more follow-up data are collected prospectively from the Taiwan Biobank. Because the cases used in adapting the PLCOT models do not overlap with the cases to be developed in the Taiwan Biobank and because the Taiwan Biobank has a large dataset, we could use datasets nonoverlapping with those used for adaptation to give an independent validation study and obtain a more conclusive comparison of PLCOT-1 and PLCOT-2; in particular, we will pay attention to the calibration and discrimination around or above the risk thresholds relevant to the LDCT lung cancer screening thresholds. Note that only 6 ever-smokers aged 50 to 74 in the Taiwan Biobank developed lung cancer based on the current follow-up dataset.

Because the smoking rate among males has been decreasing in Taiwan, especially since the 2009 implementation of the Tobacco Hazards Prevention Act (47), constantly updating risk models involving smoking exposure is especially desirable.

This study takes advantage of datasets and information from multiple sources in Taiwan, which have been shown to be valuable research resources. A particular strength of this study is that the TCR from 2011 to 2016 provided smoking status information on approximately 83% of the lung cancer patients in the TCR and that among the ever-smokers, 82% had their number of pack-years smoked and quit time available; see Tables S4A and S5A.

This paper considered an age-matched case–control study because the GELAC was initially a frequency matched design. However, it seems possible to extend the current methods to the situation where cases and controls are independently sampled from their respective populations.

The methods described in this paper could adapt other well-studied absolute risk models for different populations, which would help alleviate health disparities due to the lack of risk prediction models (15). For example, the Gail model for breast cancer among Asian American females (1, 8, 48), lung cancer risk models studied in Ten Haaf and colleagues (21) and Katki and colleagues (22), and more recent breast and lung cancer risk models that incorporated polygenic risk scores (29, 49) could be validated or adapted for Taiwan use following the same methods of this paper.

Although this paper provided information useful for designing lung cancer screening programs in Taiwan, it implicitly assumed the efficacy reported in the US National Lung Screening Trial that a reduction of 20% in lung cancer mortality was observed in the LDCT screening arm (50). Eventually, an efficacy study of LDCT screening in Taiwan is desirable, and the results of this paper might be useful in such a study.

This paper presented methodologies for recalibrating or adapting risk models by synthesizing data sources without prospective cohorts and offered preliminary information useful for policy-making on designing lung cancer screening programs in Taiwan. Further studies regarding implementation, such as cost-effectiveness, are warranted.

K.Y. Chen reports personal fees from AstraZeneca, Roche, Boehringer Ingelheim, Eli Lilly, Pfizer, Novartis, Merck Sharp & Dohme, Chugai Pharma; and personal fees from Takeda outside the submitted work. H.H. Hung reports grants from Ministry of Science and Technology during the conduct of the study. C.J. Chen reports grants from Ministry of Science and Technology during the conduct of the study. I.S. Chang reports grants from Ministry of Health and Welfare, Taiwan during the conduct of the study. C.A. Hsiung reports grants from Department of Health Taiwan, Ministry of Health and Welfare Taiwan; and grants from Ministry of Science and Technology Taiwan during the conduct of the study. No disclosures were reported by the other authors.

The funders had no role in the study design, data collection and analysis, decision to publish, or manuscript preparation.

L.-H. Chien: Conceptualization, data curation, formal analysis, investigation, methodology, writing–original draft, writing–review and editing. T.-Y. Chen: Data curation, formal analysis, investigation, writing–review and editing. C.-H. Chen: Data curation, formal analysis, investigation, writing–review and editing. K.-Y. Chen: Data curation, investigation, writing–review and editing. C.-F. Hsiao: Data curation, investigation, writing–review and editing. G.-C. Chang: Data curation, investigation, writing–review and editing. Y.-H. Tsai: Data curation, investigation, writing–review and editing. W.-C. Su: Data curation, investigation, writing–review and editing. M.-S. Huang: Data curation, investigation, writing–review and editing. Y.-M. Chen: Data curation, investigation, writing–review and editing. C.-Y. Chen: Data curation, investigation, writing–review and editing. S.-K. Liang: Data curation, investigation, writing–review and editing. C.-Y. Chen: Data curation, investigation, writing–review and editing. C.-L. Wang: Data curation, investigation, writing–review and editing. H.-H. Hung: Data curation, formal analysis, writing–review and editing. H.-F. Jiang: Data curation, formal analysis, writing–review and editing. J.-W. Hu: Data curation, writing–review and editing. N. Rothman: Conceptualization, writing–review and editing. Q. Lan: Conceptualization, writing–review and editing. T.-W. Liu: Conceptualization, resources, data curation, funding acquisition, investigation, project administration, writing–review and editing. C.-J. Chen: Conceptualization, resources, data curation, funding acquisition, investigation, project administration, writing–review and editing. P.-C. Yang: Conceptualization, resources, data curation, funding acquisition, investigation, project administration, writing– review and editing. I.-S. Chang: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, investigation, methodology, writing–original draft, project administration, writing–review and editing. C.A. Hsiung: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, investigation, methodology, writing–original draft, project administration, writing–review and editing.

Part of the data analyzed in this study was provided by and analyzed on site in the Health and Welfare Data Science Center, MOHW, Taiwan. This study was supported by the MOHW (Project grants DOH95-TD-G-111–015 (to C.A. Hsiung), DOH101-TD-PB-111-TM015 (C.A. Hsiung), DOH102-TD-PB-111-TM024 (C.A. Hsiung), MOHW103-TDU-PB-211–144003 (to C.A. Hsiung), MOHW106-TDU-B-212–144013 (to I.S. Chang), MOHW107-TDU-B-212–114026 (to I.S. Chang)), NHRI (Project grants 95A1-BSAP01–002 (to C.A. Hsiung), NHRI-PH-110-GP-04 (to C.A. Hsiung), NHRI-PH-110-GP-01 (to C.A. Hsiung), and the Ministry of Science and Technology (MOST 103–2325-B-400–023 (to C.A. Hsiung), MOST 104–2325-B-400–012 (to C.A. Hsiung), MOST 105–2325-B-400–010 (to C.A. Hsiung), MOST 109–2740-B-400–002 (to C.A. Hsiung).

The authors thank Ms. Chia-Yu Chen and Fang-Yu Tsai for technical assistance.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.