Cost-Effectiveness of Personalized Screening for Colorectal Cancer Based on Polygenic Risk and Family History

Screening has been shown to be a cost-effective method to reduce the incidence and mortality of colorectal cancer (1–4). In countries with population screening programs, screening for colorectal cancer is based on age (5), with separate screening recommendations for those with a positive family history (6). However, genetic susceptibility also plays an important role in colorectal cancer risk and it has been suggested that, when combined with family history, this may improve risk prediction and diagnosis (7, 8).

Genome-wide association studies have shown that polygenic factors, such as common, low-risk genetic variants or single-nucleotide polymorphisms (SNP), play a significant role in defining colorectal cancer risk due to their relatively high prevalence in the population (9–12). In isolation, SNPs are only weakly associated with colorectal cancer risk; however, cumulatively they explain substantial variation in risk (9, 13). A polygenic test can be used to estimate someone's polygenic risk score based on the absence or presence of specific risk alleles. Such a risk score can be used to identify individuals at several times lower and greater (0.49–3.40) colorectal cancer risk than the average population (14).

Compared with age-based screening, personalized screening provides an opportunity to stratify the population allowing screening to be tailored to an individual's risk (15). This would allow for those at lower risk to start screening later and/or have longer screening intervals, whereas those at higher risk could start screening earlier, undergo more intensive screening or both (9, 15–17). Personalized screening also provides opportunities to detect cancers in younger at-risk individuals, who are currently excluded from age-based screening despite being at increased risk (18–20). In this way, personalized screening has the potential to reduce the harms of screening while maintaining, or even increasing, its benefits in addition to improving its cost-effectiveness.

Previous research has demonstrated the efficacy of stratifying the population for screening based on age and polygenic risk (21, 22), or in combination with other factors including family history (7, 23). However, no studies have evaluated the cost-effectiveness of such risk-stratified screening compared with uniform screening for colorectal cancer. To address this gap in knowledge, we investigated the impact of personalizing colorectal cancer screening, based on polygenic risk and family history and compared its cost-effectiveness to uniform screening.

We used the Microsimulation Screening Analysis-Colon (MISCAN-Colon; ref. 24) model to estimate the costs, benefits and harms of different uniform screening strategies as well as personalized screening strategies that were based on polygenic risk and family history of colorectal cancer.

MISCAN-Colon is a well-established microsimulation model for colorectal cancer developed at the Department of Public Health, Erasmus University Medical Center (24). The structure, underlying assumptions, and data sources used to calibrate the model are detailed in Supplementary Methods and Materials. In brief, the model simulates a large population of individuals from birth to death, first without and then with screening for colorectal cancer. As each simulated person ages, one or more adenomas may arise and some can progress in size from small (<5 mm) to medium (6–9 mm) to large (>10 mm). Some adenomas develop into preclinical cancer and subsequently progress through cancer stages I to IV. During each stage symptoms may present and colorectal cancer may be diagnosed. The introduction of screening may alter the simulated life histories through detection and removal of adenomas or through detection of colorectal cancer at an earlier stage with a more favorable survival. By comparing the life histories of a simulated population being screened to the corresponding life histories in a simulated population not screened, MISCAN-Colon quantifies the effectiveness and the costs of screening.

MISCAN-Colon was calibrated to match the age-specific incidence of colorectal cancer in Australia before the introduction of biennial fecal immunochemical test (FIT) screening for those ages 50 to 74 in 2006 (25). Stage distribution, localization of cancers in the colorectum, and five-year relative survival after clinical diagnosis of a cancer were based on Australian literature (26, 27). Additional assumptions of the MISCAN-Colon model are presented in Table 1 and Supplementary Methods and Materials.

For this analysis, we simulated a cohort of 100 million 40-year-olds, with life expectancy as observed in Australia in 2013 to 2015 (28). Individuals were followed for a lifetime, until a maximum age of 100 years, at which point they are all assumed to be dead.

Using previous research, the population was stratified a priori into five risk groups based on their quintile of polygenic risk score (based on 45 SNPs shown to increase colorectal cancer risk) and their first-degree family history of colorectal cancer (Table 2; ref. 14). The expected prevalence of each of the five categories in the general population was based on a random assignment of 1,000 people given a 20% probability of being in any SNP quintile and a 10% probability of having at least one first-degree relative with colorectal cancer (29). The relative risk (RR) of developing colorectal cancer (compared with the average population risk) for each risk group was based on the combined RR of each quintile of polygenic risk score and family history based on observed virtual independence of the two factors (14). The five risk groups were defined as “very low” (RR < 0.5), “low” (RR between 0.5 and 0.9), “average” (RR between 0.9 and 1.2), “high” (RR between 1.2 and 1.8), and “very high” (RR > 1.8). We assumed no other differences in life expectancy, colorectal cancer stage distribution, survival, or screening performance characteristics between the risk groups.

In addition to a scenario without screening, we modeled 25 different screening strategies, varying screening start age (40, 46, 50, 54, or 60 years), test (FIT or colonoscopy), and interval (annual, biennial or triennial screening for FIT, and every 5 or 10 years for colonoscopy). For all FIT analyses, we assumed a positivity of 7.7% based on rates observed in the Queensland Bowel Cancer Screening Program between August 2006 and December 2010 (30, 31). Screening was always assumed to stop at age 74 years. Surveillance intervals and stop-age for all scenarios were based on the Australian National Health and Medical Research Council Clinical Practice Guidelines for Surveillance Colonoscopy (32).

Screening programs can be assessed under the assumption of perfect adherence or observed adherence. In the first analyses, we assumed perfect adherence to all screening, diagnostic, and surveillance tests. Subsequently, we estimated the costs and effects of screening at adherence levels currently observed in Australia.

For the latter analysis, we simulated participation rates as reported by the Australian National Bowel Cancer Screening Program (NBCSP), a biennial FIT screening program, in 2017 (Table 1; ref. 33). Participation with annual and triennial FIT and with primary colonoscopy screening was set at the same screening participation rates. Age-specific participation rates were provided in 5-year age intervals between 50 and 74 years. As data were not available for screening participation for individuals ages 40 to 49 years, participation was assumed to be equal to those ages 50 to 54 years. We assumed that 76.0% of individuals who had previously attended screening would attend again in the next screening round, whereas 19.7% of individuals who had not attended in the previous round would now attend based on data from the NBCSP (33).

A positive FIT requires a consultation with a primary care provider, such as a general practitioner (GP), to discuss test results and obtain a referral for colonoscopy. For the observed adherence analyses, it was assumed that 90% of FIT-positive cases would attend this appointment (26). In addition, attendance at diagnostic colonoscopy was age specific ranging from 68.2% to 72.3% based on outcomes from the NBCSP (33). The participation rate for colonoscopy surveillance was assumed to be 80% (34).

The cost of screening with FIT was based on commercially available kits (35). This cost includes the test, postage, and test processing fees. The cost to analyze a FIT specimen was based on the Australian Medicare Benefits Schedule (MBS; ref. 36). The cost of attending a GP to obtain a referral for colonoscopy (standard consult) is set in the MBS (37). The cost for colonoscopy and complications from colonoscopy was obtained from the Independent Hospital Pricing Authority report on costs in Australian public hospitals (38). Costs for cancer care were based on costs of cancer treatment in the Australian setting (39). All costs are presented in Australian dollars ($AUD), standardized to 2016 prices using the consumer price index where necessary (40).

To determine risk, we assumed all individuals underwent assessment for family history of colorectal cancer and polygenic testing prior to the commencement of colorectal cancer screening. We assumed assessment for family history of colorectal cancer would be undertaken by a GP and cost the same as a standard consult (Table 1; ref. 37). In addition, we assumed polygenic testing would cost $200 based on a commercially available polygenic test for breast cancer (41).

The assumed utility loss due to colorectal cancer screening was 0.00274 quality-adjusted life years (QALY) per colonoscopy (1.5 days at 0.5 utility) and 0.001918 QALYs per complication of colonoscopy (14 days at 0.5 utility; Table 1). We also assumed that life years (LY) with colorectal cancer care are of lower quality than those without colorectal cancer care (42). We assumed no disutility from determining or knowing polygenic risk score.

For all scenarios, the model estimated health effects such as colorectal cancer incidence and colorectal cancer mortality, and required resources such as the number of screening and surveillance tests performed between ages 40 and 74 years. From these outcomes, we calculated costs, life years, and QALYs lived with each strategy. Costs, life years and QALYs were discounted at an annual rate of 5%, as is recommended in Australia (43). Undiscounted results are presented in Supplementary Results (Supplementary Tables S6 and S7; Supplementary Fig. S8).

Our analysis consisted of four parts. First, we determined costs, benefits, and harms of the aforementioned screening strategies applied to the population as a whole (uniform screening). We plotted the uniform screening scenarios in a cost-effectiveness plane and performed an incremental cost-effectiveness analysis to see which scenarios were efficient.

Second, we followed the above steps for each risk group and used these results to determine the efficient screening strategies for each risk group. Then, we combined the efficient screening strategies for all risk groups and ordered them from least expensive to most expensive. Using this list, we developed a series of optimally personalized screening scenarios. As each personalized screening scenario can be a combination of different strategies for each risk subgroup, there will be many more personalized screening scenarios.

Third, we compared the outcomes of uniform and personalized screening to establish which method would yield better results. We did this by plotting all uniform and personalized screening scenarios in a single cost-effectiveness plane and by performing an incremental cost-effectiveness analysis to see whether personalized screening or uniform screening was most efficient.

Finally, we applied imperfect participation rates to uniform and personalized screening scenarios to determine their impact in a “real-world” scenario. The benefits and costs of screening were compared with the same population undergoing no screening.

At each step, scenarios with the highest incremental cost-effectiveness ratio (ICER) under a threshold of $50,000 per QALY gained were identified as the optimally cost-effective strategy as this is a commonly used willingness-to-pay (WTP) threshold in Australia.

In sensitivity analyses, we assessed the impact of weighting QALYs by age (ref. 44; we applied age-adjusted health-related quality of life so that quality of life deceased with increasing age) and discounting our results at 3% rather than 5% as this is a common international discounting rate (45). In addition, we explored the impact of changes in screening participation for personalized screening, holding the participation of uniform screening at current levels. To do this, we increased and decreased age-specific participation of the initial screening offer by 10 percentage points and adjusted the participation of rescreening.

Due to the uncertainty surrounding costs for determining risk profile, we also included a sensitivity analysis where these costs were excluded. Using this information, we conducted a threshold analysis to estimate the maximum cost for determining risk profile where personalized screening would be considered cost effective compared with uniform screening, at a WTP of $50,000 per QALY gained.

Compared with no screening, the uniform screening scenarios (Table 3a) reduced colorectal cancer incidence by 22% to 69% (18–58 fewer colorectal cancer cases per 1,000 individuals) and mortality by 35% to 79% (10–23 fewer colorectal cancer deaths). These scenarios yielded 0.11% to 0.32% more QALYs (20–58 additional QALYs) and costs increased by 0.5% to 424% ($6,409–$5,277,930) per 1,000 individuals. These screening scenarios increased colonoscopy demand by 383 to 6,927 colonoscopies per 1,000 individuals (Table 3a). Several uniform screening scenarios were on the efficient frontier (Supplementary Fig. S1). Using a WTP threshold of $50,000 per QALY gained, the optimal uniform screening scenario was annual FIT from 50 to 74 years (ICER $43,174). Although close to the efficient frontier, biennial FIT screening from 50 to 74 years, the screening program currently implemented in Australia, was dominated. Colonoscopy screening scenarios were the most effective; however, they also had the highest ICERs.

The efficient frontier included many of the same strategies for each risk group; however, the ICERs differed substantially (Supplementary Table S1; Supplementary Fig. S2A–S2E). For example, annual screening with FIT from 54 to 74 years was on the efficient frontier for all risk groups; however, the ICERs ranged from $86,929 for those at very low risk to $3,687 for those at very high risk. Considering a WTP threshold of $50,000 per QALY gained, the optimal screening strategy for those at very low risk was biennial FIT from 54 to 74 years (ICER $33,639), whereas for those at highest risk, the optimal strategy was 5-yearly colonoscopy from 50 to 74 years (ICER $39,568). Biennial FIT screening was only on the efficient frontier for the very low risk group (ICER $63,911).

Using these efficient strategies, 39 personalized screening scenarios were created (Table 4). These scenarios (Table 5a) reduced colorectal cancer incidence by 4% to 68% (3–57 fewer colorectal cancer cases per 1,000 individuals) and mortality by 5% to 79% (2–23 fewer deaths). In addition, they yielded 0.02% to 0.32% more QALYs (3–58 additional QALYs) and increased costs by 19% to 432% ($233,599–$5,330,249). The personalized screening scenarios increased colonoscopy demand by 45 to 6,698 colonoscopies per 1,000 individuals (Table 5a).

At a WTP threshold of $50,000 per QALY gained, the optimal personalized screening scenario consisted of the following: those at very low risk or low risk, screening should start at age 54 with biennial and annual FIT, respectively, those at average and high risk, screening should start at age 50 with annual FIT, and those at very high risk, screening should start at age 50 with 5-yearly colonoscopy (ICER $45,682).

When compared, personalized and uniform screening scenarios similarly reduced colorectal cancer incidence and mortality and yielded similar QALYs. Personalized screening more efficiently allocated colonoscopy demand; however, it cost more than uniform screening, due to the cost of determining risk. Although several scenarios from each type of screening were on the efficient frontier (Fig. 1A), all of the personalized screening scenarios had an ICER above $100,000 and would therefore not be considered cost effective. At a WTP threshold of $50,000 per QALY gained, the optimal screening scenario was annual FIT screening from 50 to 74 years.

As might be expected, the application of realistic participation rates decreased the health benefits as well as the costs of all screening scenarios. At this level of participation, none of the personalized screening scenarios were cost effective compared with uniform screening (Fig. 1B).

Our results were not sensitive to changes in discounting, weighting of QALYs or adjustments to rates of participation (Supplementary Tables S2–S5; Supplementary Figs. S3–S6). However, excluding the costs of determining polygenic risk had a significant impact on our results with personalized screening dominating uniform screening scenarios at both perfect (Fig. 1C) and realistic adherence (Supplementary Fig. S7). The threshold analysis indicated that for personalized screening to be cost effective compared with uniform screening at the WTP of $50,000 per QALY gained, the cost for determining risk should not exceed $47.52 (Supplementary Table S8).

We investigated the impact of personalizing colorectal cancer screening based on polygenic risk and family history. We found that uniform screening was equally effective (cancers and deaths prevented) but more cost effective than personalized screening. Although personalized and uniform screening showed similar reductions in colorectal cancer incidence and mortality and similar gains in QALYs, personalized screening incurred additional costs resulting from the whole population undergoing testing to determine their colorectal cancer risk.

The concept of personalized screening is promising and has previously been shown to be more effective than a strategy based on age alone (7, 21–23). Our results add support to these findings. However, our results do not align with recent findings that risk-stratified screening based on polygenic risk profile for breast cancer is cost effective compared with the standard age-based screening program (46). This discrepancy may be due to differences in the discriminatory performance of risk-stratification algorithms or differences in the cost for determining risk, which was substantially lower in this analysis (£50 or ∼$90) than in ours (∼$240). However, it is difficult to accurately determine how much of the cost for establishing risk should be allocated to a screening program. The cost of polygenic testing varies widely (47), and there is potential to combine testing for other cancers. Given this difficulty and because cost-effectiveness of personalized screening is highly dependent on these additional costs, we assessed the impact of excluding them. We found that when these costs were excluded, personalized screening was cost effective. The threshold analysis suggested that at a WTP of $50,000 per QALY gained, the cost to determine risk should not exceed ∼$48, which is significantly lower than the cost assumed in this analysis.

The effectiveness of personalized screening will be affected by the precision with which the population is stratified (15). This will be affected by both the accuracy of the metric used to stratify the population and the proportion of the population willing to undertake polygenic testing. Although our results appear unfavorable, the advantage of screening based on polygenic risk and family history remains limited largely because the current contribution of known SNPs to colorectal cancer risk is modest (9, 10, 48). As new SNPs are identified, the discriminatory utility of polygenic testing will increase and the performance of risk assessment based on this metric could improve (15). The inclusion of other factors in risk assessment, such as obesity and smoking status, may also enhance the discriminatory performance of personalized screening (13, 49, 50). It may also be pertinent to consider results from an individual's screening history. As these factors will vary over an individual's lifespan, assessment of risk may need to become more dynamic in nature, and although such inclusions will present challenges, they will likely improve the harm–benefit ratio of colorectal cancer screening.

In addition, although genetic testing for colorectal cancer has been shown to be acceptable to the community (51), individuals may not always be willing to undergo testing, for a variety of reasons, including concerns over privacy, possible misuse of data, and potential negative psychological impacts of findings (52–54). For this analysis, we assumed all individuals would undergo testing to determine their risk profile; however, due consideration of how to manage this issue is required.

The benefits of population screening are largely dependent on participation. With many countries already experiencing suboptimal levels of participation in routine age-based screening for colorectal cancer (33, 55), personalized screening presents an interesting proposition. On the one hand, increasing the complexity of screening may reduce participation in screening, thereby diminishing the modest benefits. However, individuals at increased risk of colorectal cancer have been shown to be more compliant to screening guidelines than those at average risk (56), suggesting that the provision of risk information may assist in screening uptake (16, 57). Coupled with evidence that involvement of GPs improves participation in colorectal cancer screening (58), a simple risk assessment has the potential to positively affect screening participation (59). When we applied realistic rates of participation, we found that personalized screening remained suboptimal compared with uniform screening, even when participation in personalized screening was improved (Supplementary Fig. S6). This suggests that at present, increasing participation in uniform screening will likely yield better results.

Screening effectiveness will also be affected by the choice of screening test and screening frequency. This will largely be determined by a health system's capacity to provide a given intervention to its population. Our analysis indicates that screening scenarios utilizing colonoscopy are the most effective scenarios. However, as would be expected, these scenarios significantly increase colonoscopy utilization. Although personalized screening more efficiently allocated colonoscopy utilization compared with uniform screening, such increases in demand will likely be infeasible, especially in countries with limited colonoscopy capacity.

Moving from an age-based screening program will result in a redistribution of the harms and benefits. This raises ethical issues as although personalized screening may be optimal at a population level, individuals may experience increased harms or reduced benefits as a result of their screening protocol. As would be expected, our analysis indicates that when individuals undergo less frequent screening (either by starting screening later or by having a longer screening interval), they experience higher colorectal cancer incidence and mortality. However, this will be partly offset by a reduction in other harms such as invasive tests, false-positive test results, adverse events, anxiety, and inconvenience. These concerns hold for the inclusion of younger individuals, although recent evidence suggests that their inclusion is favorable (60).

Limitations exist with our research. First, we only considered a limited number of risk categories. Effectiveness and cost-effectiveness could be further improved as the discriminatory performance of risk stratification improves. Second, we only included a limited number of low-intensity screening strategies. It is possible that other low-intensity screening strategies, such as one-off colonoscopy or less frequent FIT screening would be more efficient. Third, we did not compare the (cost-)effectiveness of stratifying the population based on family history alone. However, as the aim of this research was to explore the possible implications of combining SNPs and family history in a risk assessment, and, as determining polygenic risk is assumed to be quite expensive, such a comparison would potentially make this analysis look even less cost effective. Finally, there is some uncertainty regarding the assumptions for participation in screening. We assumed that participation in screening of any form would be equal to participation in uniform biennial FIT screening. However, this is unlikely as participation in screening varies widely (5). Unfortunately, to date, there are few data examining multiple screening modalities within one population to adequately address this concern.

In summary, this research presents an exploration of the possible impact of personalized screening for colorectal cancer based on polygenic risk and family history. Our results suggest that although personalizing screening based on colorectal cancer risk is slightly more effective than screening based on age alone, it is currently not necessarily cost effective. Cost-effectiveness of personalized screening will depend on the costs of determining risk and the magnitude of the benefits of personalization. Our analysis suggests that the currently assumed cost of determining risk is too high compared with the gains and costs must be substantially lower for personalized screening to become cost effective. The balance of cost and benefits will be contingent on the discriminatory performance of risk-stratification algorithms on polygenic risk and family history, which remains suboptimal.

However, we cannot ignore the changing landscape that advances in technology provide and, as improvements in risk stratification occur and costs for polygenic testing decrease, personalizing screening will become an increasingly cost-effective and attractive option. This consortium of researchers, and others, have previously called for the concept of personalized screening to be brought to the attention of key stakeholders (15, 61). Our research seeks to highlight the possible implications of personalized screening based on risk assessment, which we believe can and will play a significant role in improving our screening programs. As such, we reiterate our call that key stakeholders carefully consider the evidence for personalized screening in order to effectively plan for the future.

M.A. Jenkins reports receiving other commercial research support from and has ownership interest (including patents) in Genetype Pty Ltd. No potential conflicts of interest were disclosed by the other authors.

The funders of the study had no role in the study design, data collection, data analysis, data interpretation, or the writing of the report. D.R. Cenin and I. Lansdorp-Vogelaar had full access to all the data, and D.R. Cenin had final responsibility for the decision to submit for publication. The contents of the manuscript are solely the responsibility of the authors and do not necessarily represent the official views of the funding bodies.

Conception and design: D.R. Cenin, M.A. Jenkins, D.B. Preen, P.C. O'Leary, I. Lansdorp-Vogelaar

Development of methodology: D.R. Cenin, S.K. Naber, A.C. de Weerdt, P.C. O'Leary, I. Lansdorp-Vogelaar

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): D.R. Cenin, M.A. Jenkins

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): D.R. Cenin, S.K. Naber, A.C. de Weerdt, M.A. Jenkins, H.C. Ee, P.C. O'Leary, I. Lansdorp-Vogelaar

Writing, review, and/or revision of the manuscript: D.R. Cenin, S.K. Naber, M.A. Jenkins, D.B. Preen, H.C. Ee, P.C. O'Leary, I. Lansdorp-Vogelaar

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): D.R. Cenin, S.K. Naber

Study supervision: D.B. Preen, P.C. O'Leary, I. Lansdorp-Vogelaar

D.R. Cenin, D.B. Preen, H.C. Ee, P. O'Leary, and I. Lansdorp-Vogelaar received funding from the Cancer Council Western Australia Capacity and Collaboration Building Grant. This research also benefited from our participation in the NCI's Cancer Intervention and Surveillance Modeling Network (CISNET; grant number: U01CA199335). M.A. Jenkins is a senior research fellow for the Australian National Health and Medical Research Council.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.