Modelling the Economic Impact of Next Generation Sequencing and Precision Medicine on Childhood Cancer Management—a Microsimulation Approach

Precision medicine is a new approach to identify the best treatment available to patients based on their genomic information. However, no economic evaluation of genome sequencing has been reported for the treatment of childhood cancers, which is critical to evaluate the feasibility of implementing patient’s genome sequencing as part of a publicly funded treatment strategy. We have developed a microsimulation model, PeCanMOD, to evaluate the cost and benefit of applying the Next Generation Sequencing (NGS) in the management of childhood cancer. This paper describes the construction of PeCanMOD. We used linked datasets of children under 18 year of age, living in New South Wales (NSW), Australia, who have had cancer, as a base population. Their records were extracted from the NSW Central Cancer Registry and were linked to mortality and hospital datasets. In addition, we simulated the genomic landscape of the cancer registry population, through information obtained from 1,200 molecularly profiled paediatric cancer from the Foundation Medicine. The model simulated the number of individuals eligible for precision medicine, and the incremental cost of treatment per life year gained if precision medicine was introduced for late stage cancer patients as a final treatment option. Cost of drugs, and hospital admission were included in the model. Data on response rate and probability of survival was imputed based on the latest available evidence. Each unit record in the model was weighted using input from the Australian Institute of Health and Welfare (AIHW) to reflect total paediatric cancer population in Australia. The model demonstrates the application of microsimulation modelling to simulate the impacts of NGS and precision medicine on costs and health outcomes for childhood cancer. JEL classification: C1, C3, I1 DOI: https:// doi. org/ 10. 34196/ ijm. 00230


Introduction
Treatment of paediatric cancer is one of the greatest success stories of modern medicine  and the success was exemplified by the treatment of acute lymphoblastic leukaemia (ALL), one of the most common types of paediatric cancers. The disease has progressed from being incurable in medicine studies. In this paper, we have described the structure of the model, and expected outputs from PeCanMOD. To our knowledge, this is the first such model of this kind applied to precision medicine and paediatric cancer.

Base population
The model is developed with inputs from multiple datasets (Figure 1). The base population in the model is sourced from the New South Wales (NSW) Central Cancer Registry which contains records of people who have had cancer in NSW (Cancer Institute, 2018). The study cohort comprised children

Data sources
• NSW Central Cancer Registry maintains records of all cases of cancer diagnosed in NSW residents. To study paediatric cancer population, we selected data of all NSW cancer patients diagnosed with cancer under the age of 18. Key variables used in the model include linkage ID, timing of cancer diagnosis, cancer types, and cancer stages at diagnosis. • NSW Admitted Patient Data Collection includes records for all hospital separations (discharges, transfers, and deaths) from all NSW public and private hospitals and day procedure centres. Key variables used in the model include timing of each admission, cost of hospital admission, separation outcome, frequency of hospital admission, principle diagnosis and secondary diagnosis, and procedures performed.

PeCanMOD structure
Globally, there were multiple large-scale precision medicine clinical trials designed for treating highrisk paediatric cancer (Chang et al., 2016;Harttrampf et al., 2017;Khater et al., 2019;Mody et al., 2015;Wong et al., 2020;Worst et al., 2016). Patients who had previous treatment failure, experiencing cancer relapsed, or were diagnosed with high-risk cancer (less than 30% 5-year survival rate) were the major participants in these precision medicine studies.
To reflect the current practice, the model assumed that individuals who were eventually decreased due to their illness would be simulated to have been eligible for precision medicines in our simulation prior to their death. These individuals were assumed to be high risk patients, who were unlikely to be cured with the current treatment regime. We identified these individuals from the NSW Central Cancer Registry by linking the dataset to other administrative datasets. Each individual in the NSW Central Cancer Registry dataset was assigned a unique identifier by the Centre for Health Record Linkage (Centre for Health Record Linkage, 2018), and data for these individuals were extracted from the other administrative datasets such as the NSW Registry of Births, Deaths, Marriages, and NSW Cause of Death Unit Record File. Records were then linked based on the unique identifier for each patient. The NSW Registry of Births, Deaths, and Marriages death registrations and the NSW Cause of Death Unit Record File records date of death, and cause of death ( Figure 1).

Imputation of genomic variants
Understanding the prevalence of genomic variants responsible for cancer development is critical to estimate the effectiveness of a precision medicine programme (Subbiah et al., 2018). We imputed genomic variants responsible for cancers using published data from the Foundation Medicine Pediatric Portal (Chmielecki et al., 2016;Chmielecki et al., 2017). The dataset consists of the molecular profiles of over 1200 paediatric tumours sequenced by the Foundation Medicine. We estimated the prevalence of genomic variants in each cancer type from this dataset. Imputation of having specific genomic variant was carried out based on the distribution of genomic variants in each cancer type (55 categories) and allocated using Monte Carlo simulation method to the matched cancer type. We have considered other data sources (Gröbner et al., 2018;Ma et al., 2018;Rusch et al., 2018), but none were as comprehensive as the Foundation Medicine dataset in terms of cancer types ( Table 2).

Simulation
To model genomic variants, we assigned a random value between 0 to 1 drawn from a uniform distribution to individuals in the cancer registry. Controlling for cancer types, if the value falls between the Continued upper and lower bounds for a gene, then the individual would be assigned the associated gene in this simulation. The simulation process was repeated 1,000 times. The treatment protocol was mostly based on one of the largest paediatric cancer precision medicine trials, NCI-COG Pediatric MATCH (MATCH) ( Table 3) (Allen et al., 2017). Due to limited evidence of the effectiveness of precision medicine as well as limited understanding of the distribution of actionable variants within the patient population, we have made several assumptions in our current model, and the model will be updated as results from the precision medicine trials become available. Treatment response rate and survival duration used in this model were sourced from clinical trials results on adult cancers (Tables 4 and 5) as there were no reported outcomes of these medications for childhood cancer cohort. It is possible that children's response to these medicines may be different to their adult counterpart's (Joseph et al., 2015). A one-way sensitivity analysis on response rate and duration of response will be conducted to estimate resulting cost and effectiveness in best-and worst-case scenarios (range of input parameters are described in Tables 4-6). Method for sensitivity analysis is described in 2.10.
The model simulated the number of individuals eligible for precision medicine, and the cost of treatment per life year gained if precision medicine was introduced to late stage cancer patients as final treatment options ( Figure 2). The probability of responding to precision medicine for each genomic variant and duration of response were estimated based on published literature or reports from the U.S. Food and Drug Administration ( Table 4). They were used to estimate the incremental life years that would have been gained for each individual in our base population if they had had one of the 10 targeted therapies from the MATCH trial for their specific simulated genomic variant, with the assumption that the patients died if they failed to respond, and that if they did respond, the patients would only survive as long as the duration of response (Table 5).

Cost of Next-Generation Sequencing (NGS)
The model assumes that each patient receives Whole-Genome Sequencing (WGS) at a cost of AU$4,926 per cancer patient (range: AU$2,991-AU$45,333) (reported costs were inflated to 2019 values by the consumer price index in origin country, and converted to Australian dollars using Purchasing Power  Parities) ( Table 6) (Gordon et al., 2020;Schwarze et al., 2020;Schwarze et al., 2018;Weymann et al., 2017) . Note that the cost of WGS included all steps in the sequencing pathway, including the costs of bioinformatic analysis and returning results. We also model the cost and effectiveness if each patient receives targeted multi-gene panel sequencing versus WGS. The cost of targeted multi-gene panel sequencing is assumed to be AU$1,433 (range: AU$437-AU$10,178) per sample (Gordon et al., 2020;Hamblin et al., 2017;van Amerongen et al., 2016;Yu et al., 2018).

Cost of drugs and managing toxicity
For the base case, the cost of hospital admission for precision medicine was assumed to be the same as the cost of hospital admission for chemotherapy less the direct cost relating to pharmacy, and was estimated based on Australian Refined Diagnosis Related Groups version 8.0 (AR-DRGs). AR-DRGs is a classification system to classify patient hospital admissions by connecting the number and type of patients treated in a hospital (known as hospital casemix) to the resources required by the hospital. For drugs that do not require inpatient care, we will refer to outpatient service cost for chemotherapy administration. Costs of managing toxicity or adverse events from treatment were assumed to be AU$5,890 per month per person (based on our (unpublished) analysis of the NSW Admitted Patient Data Collection linked to the NSW Central Cancer Registry) ( Table 6).
All costs were presented in 2019 Australian dollars. For costs not originally reported in Australian dollars or in 2019 cost base, we inflated the reported costs to 2019 by the consumer price index in origin country and converted to Australian dollars using Purchasing Power Parities.

Health utility
Health Related Quality of Life (HRQoL) measures the impact of health states on patient's quality of life. In the context of cost-utility analysis, HRQoL is summarised into utility values ranging between Table 3. Genomic variants eligible for precision medicine and the corresponding drugs. 0 (death) and 1 (perfect health). Utility measurements allow for comparison of health outcomes across diseases as well as comparison between various health care interventions. In PeCanMOD, we impute utility based on cancer type, treatment phase, health outcome, age, and gender from published literature. A review by Tarride et al. (2010) has summarised the health utilities measured for Acute Lymphoblastic Leukaemia patients during treatment (range: 0.81-0.91), and survivors of various cancers . Yeh et al. (2016) measured and reported that health utility among childhood cancer survivors is significantly poorer than health utility for the general population. In addition, a systematic review and meta-analysis of child health utilities by Kwon et al. (2018) reported utilities for a wide range of health conditions, including cancers.

Disability adjusted life years (DALYs)
Health utility data is scarce for most childhood cancers, especially during the treatment phase.

To model health outcomes we used Disability-Adjusted Life Year (DALY). DALYs is a standard metric used to describe burden of disease. This value is calculated using the Years of Life Lost (YLL) and the Years Lived With Disability (YLD). The Global Burden of Disease Study (Global Burden of Disease Collaborative Network, 2018) reported the DALY burden due to childhood cancers (GBD 2017 Childhood Cancer Collaborators, 2019).
We attributed DALYs based on patients' response to precision medicine during the microsimulation. The model will first determine YLD with estimated duration of response to precision medicine and corresponding disability weight for treatment phase and cancer types. As we assumed that once patient will only survive as long as the duration of response, we will determine YLL based on the life expectancy at the counterfactual age at death.

Budget impact analysis
The model assigns a multiplier to individuals to reflect the number of childhood cancer patients within the Australian population. The Australian Institute of Health and Welfare (AIHW) published childhood cancer incidence between 1982-2015 (Australian Institute of Health and Welfare, 2019a; Australian Institute of Health and Welfare, 2019a), and the multipliers are assigned to the total costs incurred by each individual in the NSW Central Cancer Registry, by age group at diagnosis, cancer types, sex, and year of diagnosis.

Sensitivity and uncertainty analysis
We perform one-way sensitivity analysis to determine the parameters that have the biggest influence on the model outcome. This is achieved by changing one parameter at a time while keeping other parameters constant. The parameters of interest are varied between plausible extremes (input values range are described in Tables 4-6). Model outcome (incremental costs per life year gained) for each scenario is then compared against base case to identify the parameters that significantly affect model outcome.
We also conduct probabilistic sensitivity analysis to explore the robustness of model results on all model parameters such as costs of drugs and response rate to precision medicine using Monte Carlo simulation. We assumed lognormal distributions for cost of drugs and sequencing, binomial distributions for the response rates to drugs, and Weibull distributions for the duration of response.

Weighting
The PeCanMOD assigns weights to individuals to reflect an estimate of childhood cancers within the Australian population. The Australian Institute of Health and Welfare published national annual cancer incidence by cancer types and sex via its Australian Cancer Database (Australian Institute of Health and Welfare, 2019a;Australian Institute of Health and Welfare, 2019a). Person weight was assigned to each individual in the NSW Central Cancer Registry, controlling for year of diagnosis, sex and cancer types.

Validation
We carried out internal validation including debugging via code walk through to ensure that the model does not have obvious construction and syntax errors. We also compared model output against external data where available. Model input and output are assessed by paediatric medical oncologist (authors TOB, and TT) for face validity.
To validate whether the imputed genomic variants distribution is comparable with the Foundation Medicine Pediatric Portal data, we performed Spearman's rank correlation coefficient analysis and demonstrated that the distribution of genomic variants was significantly correlated between the reference population and NSW Central Cancer Registry simulation output (ρ=0.73, p<0.01) (Figure 3).

Ethics approval and consent to participate
Human Research Ethics approval has been obtained from the NSW Population & Health Services Research Ethics Committee (HREC/17/CIPHS/7). We have sought permission to waive consent from NSW Ministry of Health under the Health Records and Information Privacy Act 2002 (NSW).

Discussion
This paper described the development of a microsimulation model PeCanMOD developed to simulate costs and potential benefits from receiving precision medicine as the last treatment resort for childhood cancer patients. Application of microsimulation model to evaluate cost-effectiveness of intervention in adult cancers was not uncommon, for example, Petelin et al. (2019) modelled costeffectiveness in a subset of breast and ovarian cancers or Bongers et al modelled cost-effectiveness in non-small-cell lung cancer (Bongers et al., 2016;Petelin et al., 2019). Introduction of precision Note: Each column along the horizontal axis represents a gene responsible for cancers. Distribution of genomic variants were significantly correlated between the reference population and NSW Central Cancer Registry simulation output (ρ=0.73, p<0.01) (details list of genes is described in Appendix 1). medicine into children is more recent than for adults. Therefore, there is no microsimulation model for precision medicine in children cancer. Our aim is to fill this knowledge gap through the development of PeCanMOD. The model can be applied in evaluating the cost-effectiveness of multi-drug precision medicine program, and the model output can report the cost-effectiveness of individual drug candidate, and the impact of introducing drug candidate on government's budget, by taking into account of prevalence of target genomic variants in population of interest.
This study has some limitations. The NSW Admitted Hospital Data Collection dataset is limited to hospital admissions occurred in NSW. Therefore, treatments occurred beyond NSW hospital were not captured. We were unable to account for migration events after patients were diagnosed with cancer as this information were not available in the linked datasets. However, we expect only a very few cases of migration as only ~1.2% of the records did not have matching hospital admission records. The mortality data in our dataset was also limited to only patients that died in NSW due to cancer. Patients died outside of NSW were not recorded in our dataset. In a recent report by NSW Cancer Institute, retrospectively matching NSW Central Cancer Registry data to the National Death Index resulted in 0.16% additional death records (NSW Cancer Institute, 2020). Therefore, we do not expect this limitation to have material impact on the process of identifying patients eligible for precision medicine in our model.
In our model, the treatment protocol is largely based on NCI-COG Pediatric MATCH study. Due to limited evidence of the effectiveness of precision medicine as well as limited understanding of the distribution of actionable variants within the patient population, we have made several assumptions in our current model. To reflect current practice, we modelled that only patients with high-risk cancer or those that have experienced treatment failure to be eligible for precision medicine. However, it is possible that with improvement in technology and clinical implementation, precision medicine can be administered as soon as patients are diagnosed with cancers, and not limited to patients that experiencing treatment failure. The current assumption might introduce heavier weight into certain cancer types as individuals were selected for inclusion in the model based on survival outcomes (post-hoc), which is not feasible in reality.
Due to limited data availability, the model used results from adult trials to simulate drug efficacy in children. In optimal scenario, children are likely to react to medicine just as well as adults, however, it is also possible that they may metabolize certain medicines differently to adults resulting in severe adverse drug reactions and toxicity (Contopoulos-Ioannidis et al., 2010). Extrapolation of the therapeutic benefit from adults to children need to be treated with cautions (Janiaud et al., 2015). Therefore, this model will be continuously updated with new input parameters when relevant clinical trials data are reported.
We were not able to control for other variables, such as sex and age, as these data was not recorded in the Foundation Medicine dataset. Age and sex were significantly associated with cancer risk in this cohort (Stjernfelt et al., 2020), and controlling for age, sex, as well as cancer types, will greatly improve the impute estimates. Furthermore, the Foundation Medicine dataset was not designed to reflect the cancer prevalence in population, with emphasis on cancer types with low 5-year survival. We have imputed genomic variants onto Australian population based on US data, which may not be reflective of genomic diversity of Australian population as disease associated variant might differ across populations (Altshuler et al., 2015;Corona et al., 2013). However, we were unable to ascertain this without evidence from cross-population comparison of the distribution of pathogenic variants responsible for childhood cancer.
Apart from being required to have the appropriate genomic variants, patients also have to meet several other requirements, such as performance status (general well-being of the patients), in order to be eligible for precision medicine trials. In this model, we have simply assumed that all patients having the targetable genomic variants would also meet other eligibility criteria for precision medicine trials. As more clinical trials data becomes available, the model will be updated so that it remains current. As NCI-COG Pediatric MATCH have not yet published their clinical findings, we have used published drug responses in other settings (e.g. adult cancers) as our guide for response rate and survival. Key parameters, including the genomic landscape of paediatric cancer and the effectiveness of precision medicine, are influential in modelled outcomes. With increasingly widespread implementation of genomic sequencing, we will also assume that the cost of genomic analysis will decrease. At present, economic models for precision medicine suffer from the lack of "real world" clinical trial data inputs for the model (Terkola et al., 2017). The design of this model allows flexibility in modelling other treatment protocols, as well as determining the minimum effectiveness or maximum costs of treatment required to achieve cost-effective care. There are several paediatric oncology precision medicine trials ongoing, and health outcome results from these studies are highly anticipated.
However, it is also noteworthy that the majority of ongoing precision medicine clinical trials do not have a matching control population, therefore, modelling the cost-effectiveness of precision medicine programme would rely on counterfactual simulation in silica.
Furthermore, as highlighted in this modelling exercise, understanding the genomic landscape of paediatric cancer patients is crucial in determining the proportion of eligible participants for precision medicine. We believe that a registry created for cancer patients and linking details of the patients' genomic information would be very useful for future economic evaluation.  TAF1   TAL1  TBL1XR1 TBX3   TCF3   TERT   TET2   TLL2  TMEM30A TNFRSF11A TNKS   TNKS2   TP53   TRRAP   TSC1   TSC2   TSHR   TYK2   U2AF1   VHL   WHSC1   WT1   XPO1   XRCC2   ZMYM3   ZNF217 ZNF703