Modeling Thyroid Cancer Epidemiology in the US using Papillary Thyroid Carcinoma Microsimulation model (PATCAM)

Abstract

Objectives:

Thyroid cancer incidence increased over 200% from 1992 to 2018, while mortality rates had not increased proportionately. The increased incidence has been attributed primarily to the detection of subclinical disease, raising important questions related to thyroid cancer control. We developed the Papillary Thyroid Carcinoma Microsimulation model (PATCAM) to answer them including the impact of overdiagnosis on thyroid cancer incidence.

Methods:

PATCAM simulates individuals from age 15 until death in birth cohorts starting from 1975 using four inter-related components including natural history, detection, post-diagnosis, and other-cause mortality. PATCAM was built using high-quality data and calibrated against observed age-, sex- and stage-specific incidence in the US as reported by the Surveillance, Epidemiology, and End Results (SEER) database. PATCAM was validated against US thyroid cancer mortality and three active surveillance studies including the largest and longest running thyroid cancer active surveillance cohort in the world (from Japan) and two from the US.

Results:

PATCAM successfully replicated age- and stage-specific PTC incidence and mean tumor size at diagnosis and PTC mortality in the US between 1975 and 2015. PATCAM accurately predicted the proportion of tumors that grew more than 3 mm and 5 mm in 5 years and 10 years, aligning with the 95% confidence intervals of the reported rates from active surveillance studies in most cases.

Conclusions:

PATCAM successfully reproduced observed US thyroid cancer incidence and mortality over time and was externally validated. PATCAM can be used to identify factors that influence the detection of subclinical PTCs.

Precis:

This study presents the development and validation of the first simulation model replicating thyroid cancer incidence and mortality in the United States over time.

INTRODUCTION

Thyroid cancer incidence in the US increased over 200% (5.7 to 13.1 per 100,000 individuals) from 1992 to 2018 with over 60,000 new cases diagnosed annually.¹ Currently, thyroid cancer is the 13^th most common cancer overall and 6th leading incident cancer in females.^2,3 While the median age at diagnosis is 51,¹ it is the most common solid cancer in adolescents and young adults aged 15 and 39 years in the US.^4,5 Most of these cases involve papillary thyroid cancers (PTCs), which represent approximately 90% of all thyroid malignancies.² If undiscovered, most small PTCs will never cause symptoms during a person’s life-time and have little impact on their health, quality of life, or life expectancy. As a result, while PTC incidence increased dramatically over the past 30 years, mortality rates had not increased proportionately at 0.4-0.5/100,000.⁶

There are several open questions related to the optimal control of thyroid cancer. First, there is uncertainty regarding the primary factors driving dramatic increase. An acknowledged cause of the rising incidence without a corresponding increase in mortality rates is the increased diagnostic scrutiny related to the availability of imaging. The result is incidental identification of thyroid nodules, which leads to detection of a subclinical reservoir of thyroid cancer.^7,8 This may be due to the overdiagnosis, which is defined as diagnosis of a condition which would otherwise not lead to any symptoms or mortality.¹⁰ In fact, previous studies suggested that the rising thyroid cancer incidence without a corresponding increase in mortality over time indicates overdiagnosis.¹¹ It is also possible that there has been a co-occurring true increase in the occurrence of PTC due to the influence of recognized or unrecognized risk factors.⁶ Accurately estimating subclinical disease detection is essential to devise appropriate management approaches such as broadening of criteria for active surveillance eligibility. Uncertainty also surrounds decisions related to thyroid nodule biopsy and treatment for identified cancers^{12 13}

Randomized controlled trials (RCTs) provide the most reliable information for answering these and other relevant questions, but thus far no RCT has been conducted since an extended time horizon is needed to measure outcomes such as long-term recurrence. In the absence of robust RCT data, microsimulation models can help to fill the knowledge gap. To this end, this paper describes the development and calibration of the PApillary Thyroid CArcinoma Microsimulation model (PATCAM), which replicates thyroid cancer epidemiology in the U.S. over time. To our knowledge, only one other thyroid cancer simulation model has been developed to investigate treatment strategies for thyroid nodules.¹⁴ PATCAM differs from the other model in its ability to replicate temporal trends of thyroid cancer incidence and mortality and by incorporating important factors leading to observed incidence and mortality over time. PATCAM can therefore more accurately estimate the rate of subclinical detection of disease and determine the likely impact of changing healthcare practices related to referral, biopsy, and treatment on long-term outcomes.

METHODS

Development of PATCAM

The primary purpose of PATCAM is to estimate the rates of overdiagnosis by sex and age group with a particular focus on the impact of diagnostic pathways on the detection rates, which guided the model structure and details. As in other cancer models, modeling thyroid cancer inevitably involves making some assumptions, such as whether there is an increase in underlying thyroid cancer risk over time and whether the natural history of thyroid cancer differs by sex. In order to incorporate such structural assumptions into PATCAM, we followed a unique strategy in which we assembled a stakeholder advisory board composed of patients, providers, policymakers and researchers with diverse areas of relevant expertise. The modeling team regularly met with the advisory board, presented alternative approaches, and modified the model structure based on the feedback from these stakeholders. The list of the advisory board members and their area of expertise is included in Supplement Table 1.

Overview of PATCAM

PATCAM is a discrete-event system microsimulation model representing PTC-related events at the tumor and human level using a cycle time of 1 year. We include a brief description of PATCAM in this section and provide more details in the Appendix. Each individual enters the model at age 15 and is simulated until death either due to natural causes or thyroid cancer. PATCAM simulates all males and females in the US starting from 1920. After 1920, people who turn 15 in that year are added to the model and are aged according to the US life tables over time (Table 1). The population sizes over time are estimated using US Census Bureau data.¹⁵ Although the model simulates the US population starting from 1920, the first 55 years of the simulation (between 1920 and 1974) are used as the warm-up period.¹⁶ Through this, a reservoir of thyroid nodules and cancers is created to represent the prevalence of PTC in 1975. We modeled sexes differently because there is evidence that the lifetime prevalence of thyroid nodules,^17-19 thyroid ultrasound referral and biopsy patterns,²⁰ and palpability of nodules and tumors differ by sex.^21,22 Similarly, we modeled different age groups as there is evidence that the natural history,^23-32 prevalence,^18,19,33 and diagnostic pathways differ by age.³⁴ Age was stratified into four groups: 15-34, 35-49, 50-64, 65 and over.

Table 1.

Overview of the input parameters for PATCAM

Name	Description	Dependence	Data source	Value, according to the best parameter combination if calibrated
Population demographics
Population over time	Number of individuals by age and sex in the US (1920-2030)	Age and sex	US Census Data15	-
Other-cause mortality	The probability of death for individuals due to non-thyroid cancer-related causes (1959-2017)	Age and sex	Human Mortality Database41	-
Natural history
Lifetime prevalence	The probability that an individual develops a thyroid nodule throughout their lifetime	Sex	Calibrated	51.9% for females 45.2% for males
Relative distribution of initiating thyroid nodules	Age distribution of initiating thyroid nodules	Age and sex	Reiners et al. (2004)19	Appendix Table 2
Nodule type	The proportion of nodules that are assigned as PTC, non-PTC, or benign at the time of onset	Age	Calibrated	Appendix Table 3
Growth rate distribution	The distribution of PTC growth rate	None	Calibrated	Triangular distribution* (−0.195, 0.019, 0.119)
Proportion of the nonprogressive subclinical (NPS) tumors	Percentage of the PTCs that stop growing/regress after reaching a certain size	None	Calibrated	79.3%
Maximum size for NPS tumors	The size that NPS tumors stop growing/regress	None	Calibrated	2.0 cm
Proportion of the NPS tumors that are aggressive	Percentage of the NPS tumors that are aggressive	Age	Calibrated	9.6% for the 15-34 age group 4.0% for all other age groups
Aggressive growth rate distribution	The distribution of aggressive growth rate	None	Calibrated	Normal distribution (0.7, 0.1)
Maximum size for aggressive tumors	The size that aggressive tumors stop growing	None	Calibrated	1.7 cm
Proportion of the NPS tumors that regress	Percentage of the NPS tumors that regress	None	Calibrated	0%
Proportion of hyperaggressive tumors	Percentage of the non-NPS tumors that are hyperaggressive	None	Calibrated	0.3%
Maximum size for hyperaggressive tumors	The size that hyperaggressive tumors stop growing	None	Calibrated	4 cm
Hyperaggressive growth rate distribution	The distribution of hyperaggressive growth rate	None	Calibrated	Normal distribution (0.6, 0.1)
Annual increase in risk	Percentage increase in annual risk to develop thyroid cancer after 1975	None	Calibrated	1.0%
Minimum size to progress to regional stage	The minimum size for a nodule to progress to regional stage	None	Calibrated	2.0 cm
Minimum size to progress to distant stage	The minimum size for a nodule to progress to distant stage	Age	Calibrated	4.5cm for individuals over age 65 3.5 cm for individuals below age 65
Tumor detection
Probability of thyroid ultrasound-palpable	Proportion of individuals that undergo thyroid ultrasound after a palpable nodule	Tumor size, sex, and age group	KPWA data and calibrated	Appendix Table 7
Probability of thyroid ultrasound-non-palpable	Proportion of individuals that undergo ultrasound examination without a palpable nodule	Tumor size, sex, and age group	KPWA data and calibrated	Appendix Tables 5-6
Probability of detection	The probability that a tumor (either palpable or non-palpable) will be detected by ultrasound	Tumor size	Calibrated	0% for tumors smaller than 0.5 cm 95% for tumors with a size of 0.5 cm; 100% for tumors over 1 cm Linear in between
Probability of biopsy referral after ultrasound	Probability that the patient will be biopsied after a thyroid ultrasound	Tumor size	KPWA data and calibrated	0% for tumors smaller than 1 cm 100% for tumors larger than 2.7 cm Linear in between
Sensitivity of biopsy	Probability that the biopsy result of the patient will be true positive	None	Hsiao et al. (2022)40	86%
Post-diagnosis
Survival after diagnosis	The year that a patient diagnosed with PTC will die due to thyroid cancer	Age, sex, tumor stage	SEER42	Appendix Figure 4

^*When an individual is assigned a negative growth rate, the model uses zero growth rate instead.

PATCAM simulates the lifetimes of individuals through four inter-related components: (1) Natural history component, which simulates tumors from initialization until diagnosis; (2) Detection component, which represents the detection and diagnosis of PTC; (3) Post-diagnosis component, which represents the events experienced by a patient after a PTC is detected; and (4) Other-cause mortality component, which represents death due to non-thyroid cancer causes. The simulation flowchart representing the modeled events is presented in Figure 1.

Figure 1.

Incidence and mortality over time compared to SEER. Green and blue lines in these graphs represent the PTCs diagnosed through palpable and non-palpable diagnostic pathways using PATCAM, respectively. Red lines represent the total incidence predicted by PATCAM and black lines represent the observed SEER incidence with 95% confidence intervals. A) age-adjusted incidence (ages 15+) for all females; B) age-adjusted incidence (ages 15+) for all males; C) age-adjusted mortality (ages 15+) for all females; D) age-adjusted mortality (ages 15+) for all males.

Natural history component

The natural history component of PATCAM makes several assumptions. (1) All tumors are initiated at 1 mm. (2) Although individuals may have multiple thyroid tumors, only the largest tumor within an individual is simulated. (3) Tumor size is represented based on the maximum diameter in the shape of a perfect sphere. PATCAM uses tumor size to represent events related to diagnosis and treatment and is a proxy for likelihood of detection and staging.

Based on the underlying sex-specific lifetime prevalence of thyroid nodules (estimated via calibration as described below), PATCAM initializes thyroid nodules (benign or malignant) for individuals between ages 15 and 76. The rate for nodule onset at a given age is estimated by using the relative distribution of nodules from the study by Reiners et al. (2004) that conducted an asymptomatic screening of a large number of individuals in Germany and reported the age distribution for detected thyroid nodules (Appendix Table 2).¹⁹

PATCAM assigns an onset tumor to one of three categories: benign, PTC, or non-PTC based on the nodule type parameter, which depends on age, as there is evidence that younger individuals are more likely to have nodule growth than older individuals (Appendix Table 3).^30-32,35 The non-PTC nodules represent other major histologic types of malignant tumors that are not PTC including follicular, medullary, Hurthle cell, and anaplastic thyroid cancers.

While the same distribution for growth rate for PTCs is used for all individuals, each individual has a personalized growth trajectory. After a PTC is onset, the tumor is assumed to grow according to a Gompertz function, which is commonly used to successfully model the tumor growth of other solid cancers.³⁶

We make three additional important assumptions about the natural history. First, we assume that the underlying risk of developing thyroid cancer increased over time even independent of the increased imaging use for thyroid diagnosis.⁶ Second, some PTCs never progress and stop growing, thus rarely posing a risk of mortality due to their extremely limited malignancy potential; these are referred to as non-progressive subclinical (NPS) tumors. At the time of onset, a proportion of PTCs are assigned as NPS, which grow until they reach a threshold size (maxNPS) and some can even regress. NPS tumors are also used to represent the misclassification of normal lesions. A proportion of the non-NPS PTCs are aggressive, which grow at a fixed rate that is faster than normal PTCs. Third, a proportion of non-NPS PTCs are hyperaggressive, which grow more aggressively than the aggressive PTCs.

All newly onset PTCs are assigned a localized stage using Surveillance, Epidemiology, and End Results (SEER) historical staging. When a tumor reaches the minimum size to progress to regional/distant stages, the tumor progresses to regional/distant stage with a probability that depends on the individual’s tumor growth rate and tumor size.

Detection component

Cancer diagnosis is initiated by a referral to thyroid ultrasound through the following four common pathways:³⁷ (1) Palpable new mass pathway, which includes patients with palpable nodules; (2) Symptomatic pathway, which includes patients referred for ultrasound in the absence of a palpable nodule but with symptoms such as globus sensation, hoarseness, or dysphagia; (3) Incidental pathway, which includes patients who receive neck imaging for reasons unrelated to thyroid cancer; and (4) Known thyroid disease pathway, which includes patients with metabolic thyroid disease (e.g., hypothyroidism) who undergo ultrasound. We aggregated these pathways into two groups: those that undergo ultrasound for (1) palpable and (2) non-palpable nodules. If the ultrasound result is positive, the individual may be referred to biopsy based on biopsy referral patterns.

We used data from Kaiser Permanente Washington (KPWA) to estimate the input parameters related to the healthcare utilization and ultrasound referrals as no national data source comprehensively captures the diagnostic test and imaging utilization patterns in the US. KPWA data was used because a rich literature demonstrates the generalizability and dissemination potential of research in the KPWA population.^38,39 KPWA contributes data to the SEER Puget Sound site, which approximates observed national SEER trends, indicating that KPWA dataset provides a good proxy for nationwide trends in thyroid cancer (Appendix Figure 2).

The probability that an individual is referred to a thyroid ultrasound is represented by a piecewise linear function of tumor size, which is adjusted by sex and age group separately for palpable and non-palpable tumors (Appendix Tables 5 and 6). There is evidence that palpability of nodules differs between males and females, as males have thicker necks, which can impact clinical detection.^21-22 Additionally, the likelihood of undergoing thyroid ultrasound without a palpable nodule depends on calendar year based on KPWA data (Appendix Figure 3). The probability of a positive thyroid ultrasound is determined by the probability of detection parameter that depends on size, whereby increasing size increases the probability of nodule detection. (Appendix Table 7).

Once a tumor is detected and tumor size is determined, then a second parameter determines whether the individual undergoes biopsy, which depends on tumor size and biopsy sensitivity (Appendix Table 7). The biopsy sensitivity is assumed to be 86%, independent of sex and age group.⁴⁰ The model allows for measurement variability between pathologists implicitly by employing a random variable for this parameter. If the individual does not undergo biopsy or has a negative biopsy result, the individual is simulated the same way as an undetected individual.

Post-diagnosis and other-cause mortality components

Once a person has a positive biopsy result, PATCAM does not represent treatment patterns and different treatments explicitly. Instead, the year of death due to thyroid cancer is assigned based on the individual’s age group, sex, and SEER historical stage. For this purpose, 10-year cause-specific survival data by stage as reported by the SEER database is used and extrapolated to the future (Appendix Figure 4). We use US mortality data from the Human Mortality Database to assign a death year for all simulated individuals due to non-thyroid cancer causes.⁴¹ If the date of death from thyroid cancer precedes the expected date of death due to other causes for an individual, PATCAM records this death as thyroid cancer death.

Projections to the future

While PATCAM is primarily focused on replicating historical trends, for future projections, we assumed that ultrasound referral patterns remain the same as in 2019 and other-cause mortality remains the same as in 2017.

Calibration

Since PATCAM comprises a total of 47 parameters (19 related to natural history and 28 related to detection) that are not directly available from the input data, we utilized a calibration process to estimate their values (Appendix Table 8). This involved testing different parameter combinations to match calibration targets representing observed thyroid cancer epidemiology data in the US. Running PATCAM for every possible combination for these 47 parameters would be computationally infeasible; therefore, we followed a structured approach to calibration.

Our primary calibration target is the SEER incidence data obtained from the SEER Research Data (9 Registries Nov 2019 Sub) for the years 1975-2017.⁴² In addition to the SEER incidence targets, we also utilized secondary calibration targets, such as mean size at detection and stage-specific incidence. Since PATCAM uses tumor size to represent thyroid-cancer related events, we used the secondary calibration targets to ensure that the model’s tumor size representation and implied stage-specific incidence are consistent with reported values.

We first selected a subset of the parameter combinations using a grid-based search and a random search and complemented these using manual selection. We then used PATCAM to evaluate the SEER incidence generated by using the selected parameter combinations in PATCAM. We defined a calibration score, which indicates the predictive accuracy of each parameter combination. The calibration score is calculated by counting the number of years where predicted incidence falls within the 95% confidence intervals (CIs) of reported SEER incidence in each year. A higher score represents more accurate parameter combinations. For example, between 1975 and 2015, there are a total of 41 values for SEER incidence for each age and sex group. Therefore, a theoretical perfect parameter combination would have a calibration score of 328 (41*8) as there are a total of eight sex/age groups.

We conducted calibration in two steps recognizing that, prior to 1990, thyroid ultrasound referrals for non-palpable nodules were uncommon and SEER thyroid cancer incidence was stable.⁴² In the first step, 1975-1990 SEER data were used to calibrate only natural history and detection parameters that do not change by time, such as sex- and age-group specific probability of palpability of nodules (Appendix Tables 5 and 6). A total of 35 parameters were calibrated in this first step. In the second step, we used 1991-2015 SEER data to calibrate detection parameters that changed over time without modifying any calibrated parameters from the first step (Appendix Tables 5 and 6).

For both calibrations, we developed a neural network model to determine promising parameter combinations. Neural network are computational models inspired by human brain, with artificial neurons mirroring biological brain cells.⁴³ Neurons are grouped into layers, consisting of an input layer, an output layer, and potentially multiple hidden layers. Neural networks can automatically learn and extract features from raw data. This makes neural network capable to detect subtle and nonlinear relationship in complex dataset like our model calibration. We preferred neural networks over other models as they have been previously used successfully for calibration in other cancer simulation models.^44,45 In particular, we used bagging neural networks with 10 estimators, where each estimator is a 4-layer neural network with layer sizes of 20, 25, 40, and 15.^44,46 After training the bagging neural networks with the selected subset of parameter combinations and the corresponding calibration scores, we expanded the parameter space to search for additional promising parameter combinations, and used PATCAM to determine their fit to the calibration targets.

Validation

We achieved face validity by obtaining advisory board input for model assumptions, and then checked the final model by presenting the final model structure, assumptions, and predictions to the stakeholder advisory board. We also conducted an internal validation experiment that compared the mortality projections of PATCAM to the observed mortality in the US.

In addition, we conducted external cross-validation experiments that compared predictions made by PATCAM to active surveillance studies in which patients diagnosed with PTC underwent regularly scheduled ultrasound exams to monitor growth in tumor size over time. First, we compared the model results to data from Kuma Hospital in Kobe, Japan, site of the longest running and largest active surveillance cohort in the world.⁴⁷ Next, we used data from two US active surveillance studies: one from Memorial Sloan Kettering Cancer Center, New York, New York,²³ and one from Cedars-Sinai Hospital, Los Angeles, California.⁴⁸ In each of these external validation experiments, no modifications to the PATCAM model components and parameters were made except the following three inputs, which were matched to the reported values: 1) mean age; 2) sex distribution; 3) initial tumor size. We focused on the cumulative incidence of the tumors that grew more than 3 mm and 5 mm in 5 years and 10 years, as reported by these studies. We considered two scenarios in interpreting the patients who underwent surgery and were lost to follow-up in Kuma data. While the high-end estimate assumed that tumors of these patients grew more than 3 mm, the conservative estimate assumed that only surgeries that were performed for tumors explicitly coded to grow more than 3 mm growth were considered. Since we had access to primary data from Kuma Hospital, we plotted the cumulative incidence over time for the Kuma cohort for both scenarios. More details on the validation are presented in Appendix B.3.

Technical implementation

We implemented PATCAM using Python 3.8, supplemented with the NumPy library. For graphical representation, we used the Matplotlib library, and for neural networks, we used the sklearn library. To conduct a large number of replications in a short time, the model was executed on a high-throughput computing system that utilizes thousands of processors for computations at our institution. Due to the memory constraints, each model replication simulates one individual for every 150 Americans. There are 75 replications in each run, so a single run of the model effectively simulates half of the US population.

RESULTS

In the first step of calibration prior to 1990, our experiment generated 20 acceptable parameter combinations with calibration scores ranging between 239 and 286. The values of these acceptable parameter combinations are presented in Appendix Table 9. While the calibration resulted in a list of acceptable parameter combinations representing the heterogeneity and uncertainty in the natural history of PTC, the parameter combination that provided the best match to the calibration targets is identified as the best parameter combination, which had a calibration score of 286. Using the best parameter combination, PATCAM was able to replicate sex- and age-specific PTC incidence between 1991 and 2015 as reported by SEER (Figure 1 and Appendix Figure 6). In particular, PATCAM was able to mimic the dramatic increase in PTC incidence between 1995 and 2010 and leveling off in incidence after 2010. PATCAM predicted that if the referral patterns and diagnostic processes in thyroid cancer beyond 2018 remain similar to those between 2010 and 2018, the incidence and mortality of PTC is expected to remain relatively stable until 2030. The parameter combination that provided the best match to calibration target estimated that the proportion of NPS tumors is 79.3%, suggesting a large reservoir of nodules that may lead to overdiagnosis.

Mean size at diagnosis and the stage distribution by PATCAM approximated the reported rates from SEER (Appendix Figures 5 and 7). The difference between predicted and SEER-reported mean tumor size at diagnosis is larger in females compared to males indicating a poorer fit for females than males. PATCAM also predicted a slight reduction in mean tumor size at diagnosis, which is consistent with SEER data. In addition, PATCAM’s predicted stage distribution compared well against SEER-reported stage distribution in earlier time periods (1975-1990 and 1991-2000) whereas the fit is relatively worse for the later years (2001-2010) (Appendix Figures 5 and 8).

PATCAM successfully replicated the age-adjusted mortality by sex in the US over time (Figure 1). In addition, PATCAM accurately predicted the proportion of tumors that grew more than 3 mm and 5 mm in 5 years and 10 years, aligning with the 95% CIs of the reported rates from active surveillance studies in most cases. (Appendix Figure 9 and Table 2).

Table 2.

Validation results

Data source	Validation target	Reported value in % (95% CI)	Predicted value by PATCAM
Cedars-Sinai Medical Center	Cumulative incidence of tumors with >3 mm growth in 5 years	13.4 (6.0-23.7)	20.6
Cedars-Sinai Medical Center	Cumulative incidence of tumors with >5 mm growth in 5 years	10.8 (3.0-24.2)	11.4
Memorial-Sloan Kettering Cancer Center	Cumulative incidence of tumors with >3 mm growth in 5 years	12.1 (2.7-21.8)*	19.2
Kuma Hospital	Cumulative incidence of tumors with >3 mm growth in 5 years	High-end estimate: 10.1 (8.9-11.5) Conservative estimate: 6.9 (5.9-8.1)	18.3
Kuma Hospital	Cumulative incidence of tumors with >3 mm growth in 10 years	High-end estimate: 23.2 (20.6-26.2) Conservative estimate: 19.8 (17.2-22.8)	23.9

^*calculated using digitized curves from published cumulative incidence data.

DISCUSSION

The dramatic increase in thyroid cancer incidence between 1990 and the early 2010’s, without a proportional change in mortality requires a careful investigation. There are two main hypotheses, which may both be true to some degree. The first is that the change in incidence is attributable to a true increase in the underlying risk. The second is that the change in incidence is attributable to factors that affect rates of detection of subclinical disease, such as increased use of imaging and other changes in clinical practice patterns. To explore these questions, we developed PATCAM, the first microsimulation model that simulates thyroid cancer incidence and mortality over time while accounting for selected major known factors that contribute to the detection of thyroid cancer including the use of thyroid ultrasounds and biopsy over time by age and sex.

Internal and external validity

PATCAM successfully replicated the observed US incidence and mortality as reported by SEER over time, indicating internal validity of PATCAM. We externally validated PATCAM by comparing its predictions of the proportion of tumors that grew more than 3-5 mm to those in the active surveillance cohorts in 5 and 10 years.

Comparative models

While other common cancers (e.g., breast, prostate, colon, lung) have had multiple models built predicting observed incidence and mortality over time, PATCAM is the first thyroid cancer model to use this approach.^49,50 Thus, direct comparisons of our model to previous literature is limited. We are aware of only one other model for thyroid cancer, and it focuses primarily on incorporating biomarkers to personalize thyroid cancer treatment.¹⁴ In contrast, our model focuses on thyroid cancer at the population level, estimating overdetection rates and determining the impact of changing healthcare practice patterns and referrals on long-term outcomes of thyroid cancer. Our model has a detailed representation of the diagnosis process and was built using rich historical data.

Stakeholder role in model development

A unique aspect of the PATCAM development is that it was built with systematic input from a diverse set of stakeholders with expertise in multiple specialties and interests that contribute to thyroid cancer detection, care, and guideline publication. We identified members as specific individuals who were involved in guideline organizations and those likely to effectively disseminate the model within their respective fields and groups. All meetings included education and elicitation portions, in which the group engaged in discourse and provided feedback on key inputs and assumptions in the model. Advisory board members were surveyed on key questions on modeling assumptions. We also invited individual members to provide feedback on specific questions within their area of expertise. This approach helped us to refine the model structure and ensure the assumptions were consistent with clinical intuition.

Limitations

First, we did not differentiate the natural history of PTC regarding tumor growth between sexes whereas it is possible that biological differences exist. We did not attempt to calibrate tumor growth parameters by sex since there is no available data to create an accurate sex-specific natural history model and to prevent overfitting of the model. Moreover, while machine learning-based calibration approaches are increasingly being used by cancer modelers who rely on rigorous approaches to identify the promising parameter combinations, we did not completely rely only on machine learning to identify the most promising parameter combinations.^44,51 Instead, we combined a trial-and-error method with grid search to generate a set of initial parameter combinations for training the machine learning algorithm.

Conclusions

Overall, PATCAM effectively replicates thyroid cancer incidence and mortality over time while accounting for the evolving use of thyroid ultrasounds and biopsy by age and sex. Extensive calibration and validation experiments position PATCAM well to estimate rates of thyroid cancer overdiagnosis and to conduct experiments evaluating the effect of reducing rates of diagnostic testing on mortality. Further development of PATCAM will include quality of life and treatment components to address unique issues related to thyroid cancer survivorship. In particular, we plan to expand PATCAM so that it explicitly represents different treatment options after diagnosis, benefits and harms of different treatments such as total thyroidectomy and lobectomy, possibility of recurrence, and quality-of-life associated with events experienced by patients during survivorship.

Highlights:

• This study used high-quality data sources and rigorous calibration to develop the first cross-population microsimulation model to replicate thyroid cancer epidemiology in the US over time.

• This study demonstrated the use of a stakeholder advisory board composed of patients, providers, policymakers and researchers in building an epidemiologic microsimulation model.

• This study successfully explained the drastic increase in the incidence of thyroid cancer since 1990s through modeling of changing diagnostic pathways over time.