Evaluating the impact of multicancer early detection testing on health and economic outcomes: Toward a decision modeling strategy
We have benefitted along the way from comments and recommendations from members of the Universal Cancer Screening Summit leadership team as well as from 2 anonymous reviewers and the journal's editors.
The authors thank the Summit Participants, Planning Committee, and Host Organizations for defining a collaborative roadmap for the development, implementation, and adoption of a multicancer early detection (MCED) assay.
Emerging data provide initial support for the concept that a single, minimally invasive liquid biopsy test, performed in conjunction with confirmatory radiologic or other diagnostic testing, when indicated, could be deployed on a broad scale to screen individuals for multiple types of cancer. Ideally, such a test could do this in a way that yields a clinically important percentage of true-positive indications of cancer while minimizing false-positive signals. Modern decision modeling approaches can and should be deployed to investigate the health and economic consequences of such multicancer early detection (MCED) testing within defined at-risk populations. In this paper, through small-scale analyses involving 3 hypothetical MCED-detectible cancers, the authors illustrate the potential for MCED testing to be cost-effective, along with the pivotal role of test-induced stage shift on results. The time is ripe for additional, prospective investigations of the clinical value of MCED testing, the benefits versus the risks for screened populations, and the overall projected impact on health outcomes and costs over time.
Recent large-scale studies1, 2 provide initial support for the concept that a single, minimally invasive liquid biopsy test, performed in conjunction with confirmatory radiologic or other diagnostic testing, when indicated, could be deployed on a broad scale to screen individuals for multiple types of cancer. Ideally, such a test could do this in a way that yields a clinically important percentage of true-positive indications of cancer while minimizing false-positive signals.
The central idea, as Ahlquist noted, is that “The shared biology of tumor marker release into a common medium, like blood, can be exploited for multicancer detection from a single test.”3 Currently, guideline-embraced approaches for screening and early detection are available for only a few (although high-prevalence) cancers: breast, cervical, colorectal, lung, and prostate. A major feature of what Ahlquist originally termed universal cancer screening, and referred to herein as multicancer early detection (MCED), is the intent to pick up biologic signals from cancers that are not currently subject to routine screening. This may be because the cancer has relatively low prevalence in the general population or because there is an absence of detection strategies regarded as feasible or cost effective.
To this end, impressive early progress has been reported for emerging MCED tests. In the prospective, interventional DETECT-A study,1 which enrolled roughly 10,000 women aged 65 to 75 years who had no reported history of cancer, investigators demonstrated that a multianalyte blood test incorporating DNA and protein biomarkers—accompanied by positron emission tomography-computed tomography (PET-CT) screening for positive tests—could detect various cancers with high specificity (>99%) and overall sensitivity approaching 30%. There was accompanying evidence of the technical feasibility and acceptability to patients of such testing. The prospective case-control Circulating Cell-free Genome Atlas (CCGA) study2—involving deidentified biospecimens from greater than 15,000 patients with cancer and controls across 142 North American sites—examined whether targeted methylation analysis of circulating cell-free DNA could detect the presence of cancer and its clinical location (tissue of origin). Investigators reported that such a blood test could identify and localize cancers with >99% specificity, although sensitivity differed notably across cancers and by disease stage for any given cancer type. Taken together, published studies indicate that the stage is well set for additional, prospective investigations of the clinical value of MCED testing, the benefits compared with the risks for the screened population, and the overall projected impact on health outcomes and costs over time. These important considerations provide the launch point for this report.
A primary aim for candidate MCED tests is to identify cancer at an earlier disease stage than would have been the case through the traditional symptom-driven detection and diagnosis approach. If the MCED test can increase the proportion of individuals diagnosed with early stage disease, then a highly anticipated outcome—consistent with population-based cancer surveillance data—is that cancer-specific mortality would decline.3 And with such a decline, other desirable health-related outcomes would follow, including increased life expectancy and quality-adjusted life expectancy for those diagnosed with cancer. Treatment costs for cancers detected earlier are generally lower than for those detected at later stages. That said, whether early stage detection would lead to lower total medical care costs over the individual's remaining life course is unclear a priori, given the resulting projected increase in life expectancy and the array of competing (noncancer) health risks faced over these additional years. All this points to the central analytical question examined here: Taking relevant health outcome and cost factors into consideration, how can one go about determining whether MCED testing represents a cost-effective strategy?
At this formative stage of MCED development, data remain limited with respect to the capability of such a test to identify cancers at an earlier stage—and hence to alter clinical outcomes. Understandably, there is little information yet on the implications for health resource utilization and costs. No doubt, such data will emerge over time as major investigators in this arena continue to conduct large-sample, field-based studies.
In the meantime, we provide and implement a framework to illustrate how decision modeling strategies, using available data, could be deployed to predict the likelihood of MCED finding cancer at earlier stages and the resulting effects on quality-adjusted life-years (QALYs) and medical care costs. On the basis of these simulated outcomes, one can proceed to a standard cost-effectiveness analysis, asking whether the incremental gain in QALYs from adopting MCED testing is worth the incremental cost of doing so, using current benchmarks about what decision-makers would be willing to pay per QALY gained. It should be emphasized that these illustrative analyses are not intended to provide a definitive assessment of the effectiveness or cost effectiveness of MCED testing in practice. Such conclusions cannot be reliably drawn from evidence and experiences reported to date nor from our modeling analyses here, which use hypothetical data and simplifying assumptions to illustrate many of the analytical issues at the core of judging whether MCED testing is cost effective.
In that regard, a central aim of this report is to stimulate further discussion and investigation about the development of full-scale, methodologically sound, and empirically well grounded assessments of MCED testing within defined populations. Any such comprehensive evaluation will likely require application of simulation modeling approaches that can portray the dynamic unfolding of cancer incidence, diagnosis, treatment, and outcomes within a defined population over time—thus investigating how MCED could favorably disrupt the flow of events. Such future analyses will likely require bringing together an ensemble of such cancer disease-site simulation models for the singular purpose of evaluating the potential benefits and costs flowing from MCED testing.
Assuming that MCED testing does become broadly available, whether to undergo screening will almost certainly be an individual-level decision, substantially influenced by provider recommendations and multiple other factors. Consequently, our decision modeling analyses here are focused directly on a specific, statistically defined individual at risk for several cancers at a particular point in time (t1). The central question becomes: is it cost effective for the individual to receive MCED testing at t1 compared with no testing at that time point?
In parallel, it will also prove useful from a computational standpoint to view this individual as a member of a large cohort of statistically identical individuals, all of whom are at risk for the same cancers at time t1. Any such individual, and corresponding cohort, could be defined in terms of a particular set of clinical, demographic, and other risk factors (eg, Hispanic males aged 60 years with no family history of cancer).
- The cohort is defined simply as 100,000 individuals aged 50 years who either have been diagnosed with or remain at risk for a specific set of cancers regarded as potentially MCED-detectible at time t1, which is also the moment that everyone turns age 50. By implication, the decision model will be built for a single member of this cohort, and MCED testing will be cost effective for the cohort if, and only if, it is cost effective for the individual.
- For this cohort, there are 3 hypothetical cancers regarded as MCED-detectible—labeled Cancer 1, Cancer 2, and Cancer 3.
- For each cancer, there are 2 defined clinical stages—the Earlier Stage (E) and the Later Stage (L). For MCED testing to lead to a stage shift, it must increase the probability of detection at the Earlier Stage (compared with no MCED testing). By implication, these analyses do not incorporate the possibility that MCED testing could detect cancer at a preinvasive stage.
- The execution of any cancer testing strategy can be defined in terms of eligibility criteria (who is to be tested), starting time (at what age does testing begin), screening interval (how frequently should testing occur), and stopping criteria (when should testing cease, based on age or other criteria). In this regard:
- ◦ All members of our age-50 cohort who have not already been diagnosed with 1 of the 3 cancers are assumed to be clinically eligible for MCED testing;
- ◦ The starting time for testing is in fact age 50; and
- ◦ Most notably, the calculations assume that, if MCED testing is undertaken, it is a 1-time screening of the cohort at age 50 (time point t1).
Clearly, if MCED testing is broadly adopted in the years ahead, it is likely that subsequent guidelines will recommend testing at specific intervals (eg, every 2 years) rather than a 1-shot screen. To analyze the cost effectiveness of MCED testing in a way that allows for periodic sequential screening of a cohort (as well as multiple different cohorts) will require innovative application of cancer site-specific simulation modeling, as discussed below. Fortunately, many of the central methodological and empirical questions at issue in assessing the cost effectiveness of MCED testing emerge, and transparently so, even under the assumption of 1-time screening.
As described below, if a cohort member has a cancer (either Cancer 1, Cancer 2, or Cancer 3) and undergoes MCED testing, there are several possible clinical outcomes: the cancer can be detected by MCED testing (at time t1); or the testing strategy can yield a false negative but the cancer will be symptom-detected upstream (defined here as within 2 years of t1), at the Earlier Stage or the Later Stage; or the cancer will be symptom-detected downstream (at some point beyond 2 years after t1), at the Earlier Stage or the Later Stage. If a cohort member with 1 of the cancers does not undergo MCED testing, that cancer will either be symptom-detected upstream at the Earlier Stage or Later Stage, or detected downstream at the Earlier Stage or Later Stage. The idea is to acknowledge, albeit in a highly simplified way, that cancers not screen-detected at t1 will eventually emerge according to some time-to-detection distribution that must somehow be specified for the decision model calculations to go forward.
Whenever a cancer transitions from Earlier Stage to Later Stage we assume it will be (almost immediately) symptom-detected. One important implication for the analysis is that only the Earlier Stage cases of the 3 cancers are available at t1 for MCED testing. Another implication, which (if anything) leans against a favorable cost-effectiveness finding for MCED testing, is that there is no consideration here of the possible benefits of screen-detecting a Later Stage cancer sooner than later. That is, the analysis is structured to focus expressly on whether MCED testing can induce a stage shift in detection, from Later to Earlier, with the attendant implications for cost effectiveness.
Figure 1 provides a high-level view of the decision model for MCED testing. In what follows below, the model is developed in considerable detail, both structurally and in terms of parameter specification.
Structuring the Decision Model and Defining the Key Parameters
- Px,E, the point prevalence rate (effectively, the probability) of an undetected Earlier Stage case of Cancer x at time t1;
- TPRx,E, the true-positive rate (test sensitivity) = P(MCED Test+ | Cancer x at Earlier Stage);
- FNRxE, the false-negative rate of MCED for Cancer x at Earlier stage = 1 − TPRx,E;
- TNR, the true-negative rate (test specificity) = P(MCED Test− | no cancer at Earlier Stage); and
- FPR, the false-positive rate of the MCED test = 1 − TNR.
- P(Test+), the probability that the MCED test will be positive, with P(Test−) = 1 − P(Test+);
- PPVx,E, the positive predictive value of the MCED test = P(Cancer x at Earlier Stage | MCED Test+);
- 1 − Σ PPV = P(no cancer at Earlier Stage | MCED Test+);
- NPV, the negative predictive value of the MCED test = P(no cancer at Earlier Stage | MCED Test−); and
- Px,E | Test− = P(Cancer x at Earlier Stage | MCED Test−).
With these parameters defined, we proceed to Figure 3, which displays the basic decision model for investigating whether MCED testing is cost effective.
- P(Det x at E by t2 | Test−) as the probability Cancer x will be detected (via symptoms) at the Earlier Stage within the interval [t1, t2], given a negative MCED test result;
- P(Det x at L by t2 | Test−) as the probability Cancer x will be detected (via symptoms) at the Later Stage within the interval [t1, t2], given a negative MCED test result; and
- P(Det x E/L after t2 | Test−) as the probability Cancer x will be detected (via symptoms) either at the Earlier Stage or at the Later Stage (E/L) at some time point beyond [t1, t2], given a negative MCED test result.
For a case of Cancer x detected upstream, it is assumed for computational simplicity that detection occurs at the midpoint of the 2-year [t1, t2] interval.
Conditional on Cancer x being detected after t2, the probability it will be symptom-detected at the Earlier Stage is denoted by P(Det x at E after t2 | Test−). The probability of symptom detection at the Later Stage is P(Det x at L after t2 | Test−). These stage-specific downstream probabilities are not shown explicitly in Figure 3 but are deployed in calculating the QALY and cost outcome values associated with Cancer x being detected after t2 (as described in detail in the Supporting Materials, Section [iv]). Specific assumptions are required about when these downstream cancers are detected. As discussed below, the (highly) simplifying assumption here is that downstream cases of Cancer x are detected at t2 + 1 = t1 + 3; that is, 3 years after the false-negative MCED test.
- P(Det x at E by t2) as the probability cancer x will be detected (via symptoms) at the Earlier Stage within the interval [t1, t2], given no MCED testing;
- P(Det x at L by t2) as the probability Cancer x will be detected (via symptoms) at the Later Stage within the interval [t1, t2], given no MCED testing; and
- P(Det x E/L after t2) as the probability Cancer x will be detected (via symptoms) either at the Earlier Stage or at the Later Stage at some time point beyond [t1, t2], given no MCED testing.
If a case of Cancer x is detected after t2, the probability it will be symptom-detected at the Earlier Stage is P(Det x at E after t2). The probability of symptom detection at the Later Stage is P(Det x at L after t2). Although these 2 parameters are not displayed in Figure 3, they play a direct role in the computing of QALY and cost values associated with Cancer x being detected after t2 (see Supporting Materials, Section [iv]).
- Qk, present value of the QALYs associated with pathway k, where all QALY flows are discounted to present value at a real (inflation-adjusted) riskless annual rate of r; and
- Ck, present value of total direct medical cost (cancer-related plus noncancer-related) associated with pathway k, where all cost flows are in US dollars (USD) and are discounted to present value at the same rate as QALYs.
Defined this way, the outcomes are consistent with what has been termed the health care system perspective by the US-based Second Panel on Cost-Effectiveness in Health and Medicine.4
With all parameters now defined, to determine whether it is cost effective for a cohort member to undergo MCED testing, one folds back the decision tree by computing the expected value of discounted QALYs, assuming testing and then no testing, and the expected value of direct medical cost, assuming testing and then no testing. From there, the incremental cost-effectiveness ratio (ICER) associated with MCED testing is calculated as ICERMCED_Test = [E(Cost | MCED) − E(Cost | No MCED)] / [E(QALYs | MCED) − E(QALYs | No MCED)] = Δ E(Cost) / Δ E(QALYs). The formulas for computing the 4 component elements above, each expressed in terms of the decision model parameters in Figure 3, are provided in the Supporting Materials, Section (i); the assumptions and procedures for discounting QALYs and cost to present value (and at an assumed annual rate of 3%) are summarized in the Supporting Materials, Section (ii).
This ICER, indicating the increase in direct medical cost incurred in producing a 1-unit increase in QALYs, is compared with the decision maker's assumed willingness-to-pay threshold for a QALY, termed λ. If ICERMCED_Test < λ, then MCED testing is regarded as cost effective.
Finally, the extent of stage shift in the detection of Cancer x associated with MCED testing is defined as Stage Shiftx = P(Earlier Stagex | MCED) − P(Earlier Stagex | No MCED). When Stage Shiftx, as computed for the representative cohort member portrayed in the decision model, is multiplied by the total number of cohort members eligible for and undergoing MCED testing, the result is the expected difference in the number of cases of Cancer x within the cohort detected at the Earlier Stage as a result of MCED testing (see Supporting Materials, Section [iii]).
- To choose MCED testing is tantamount to adopting a strategy in which any positive MCED (blood) test triggers a confirmatory positron emission tomography/computed tomography (PET/CT) scan, as described in Lennon et al.1 The MCED test sensitivity and test specificity values used in our calculations reflect the end results of this testing strategy. Moreover, it is assumed that a confirming PET/CT scan identifies the cancer disease site (Cancer 1, 2, or 3) correctly; that is, the tissue of origin is pinpointed without error. This and other assumptions related to PET/CT confirmatory testing can, and should, be critically examined in more detailed modeling analyses, as discussed later.
- All of the parameters in Figure 3 are regarded as fixed values, and the ICER is computed as if they are all known with certainty. In response, we conduct illustrative sensitivity analyses, systematically varying some key parameters (1 at a time) and examining the impact on the ICER. In a full-scale, simulation model-based investigation of MCED testing, probabilistic sensitivity analyses could be conducted.
- If a cancer is detected for the individual, whether by MCED testing or via symptoms, the QALY and cost calculations here do not explicitly incorporate the possibility of a downstream recurrence of this cancer or the emergence of a second primary cancer. Rather, the likelihood of such events and the implications for expected QALYs and cost are assumed to be woven implicitly into the outcome calculations for the decision tree defined in Figure 3, as then implemented in Figure 5. Similarly, the presence of noncancer-related competing risks are intended to be reflected in our QALY and cost outcome calculations; but we simplify the tree by implicitly assuming that, if the individual does have Cancer 1, 2, or 3, it will be screen-detected or symptom-detected before death from a competing risk. Again, such assumptions can readily be weakened in full-scale simulation models.
Specifying the Parameter Values
To proceed, base-case values must be selected for all of the parameters in Figures 2 and 3 that, collectively, drive the health and cost outcomes associated with MCED testing for the prototypical cohort member, thereby allowing estimation of cost effectiveness.
We emphasize that each of these 3 hypothetical cancers is not to be regarded as a single, clinically well defined type of tumor (eg, breast cancer or colorectal cancer) but, rather, as representative of a cluster of tumor types that share certain key characteristics or attributes. Thus, Cancers 1 and 2 are similar in that neither is subject to cancer-specific, guideline-recommended screening; but it will be assumed that these malignancies differ markedly from each other in many other respects, including the QALY and cost implications of detection at an Earlier Stage. Cancer 3, conversely, represents cancers for which there is broadly available and guideline-recommended screening. Taken together, Cancers 1, 2, and 3 are intended to encompass all of the cancers that are MCED-detectible in this cohort at time t1. Aiming the analysis at 3 (and not a possibly more realistic number like 30) cancers keeps the calculations and graphical presentations tractable, allowing for a transparent examination of connections between stage shift, clinical outcomes, and cost effectiveness.
Below are the specific parameter values characterizing the 3 modeled cancers, starting with point prevalence estimates and the assumed performance characteristics of MCED testing.
Informed by data reported by Lennon et al1 and additional working assumptions, the point prevalence rate across Cancers 1, 2, and 3 combined at time t1—counting both already detected and undetected cases—is assumed to be 1.5%. This total prevalence rate is split equally across the cancers, leading to 0.5% for each of Cancer 1, 2, and 3; then, we further assume that the percentage of prevalent cancers already detected within each cancer type is 65%, 55%, and 75%, respectively. Hence, the corresponding percentage of prevalent cancers undetected at t1 is 35%, 45%, and 25%, respectively. Setting the percentage rate for symptom-detected Cancer 3 at 75% is to reflect the impact of ongoing, guideline-based screening for this cancer, with the implication that only 25% of true prevalent cases of Cancer 3 in the cohort remain undetected at t1.
With respect to this cohort of 100,000 statistically similar individuals at t1, these assumptions together suggest a total of 500 prevalent cases of each cancer type, with the total number of undetected cases of Cancer 1, 2, and 3 being 175, 225, and 125, respectively. By implication, 100,000 − 500 − 500 − 500 = 98,500 individuals are cancer-free at t1. Hence, the point prevalence rate of undetected cases of Cancer 1 at t1 is 175 / (98500 + 175 + 225 + 125) = (175 / 99025) = 0.001767. Similarly, for Cancers 2 and 3, the prevalence rates are 0.002272 and 0.001262, respectively, whereas PNC = 98,500 / 99025 = 0.994699.
In general, an inherent challenge here is that the prevalence rate of true but undetected cases at t1 is not directly observable in the general population (in the absence of special screening studies). Based in part on cancer prevalence data from the National Cancer Institute's CanQuest data system,5 we believe the Px,E values here may be conservative within the context of our decision modeling set-up. This would suggest that the analyses lean against a finding that MCED testing is cost effective; that is, all else being equal, it is likely that the higher the rate of detectible cases in a defined population, the lower the ICER (which is consistent with a sensitivity analysis described below).
TPRx,E and TNR
The selection of sensitivity rates and test specificity for our MCED testing strategy was broadly framed by the analyses reported by Liu et al2 and Lennon et al,1 while also reflecting a desire to have these true-positive rates vary across cancers. Given the assumption here that MCED testing is capable of detecting cancers at the Earlier Stage (where there is presumed to be some amount of informative molecular and biologic signaling) and not at a preinvasive stage, test sensitivities have been set at 30% for Cancer 1, 70% for Cancer 2, and 50% for Cancer 3. Here and elsewhere below, we have attempted to select a plausible centered value, namely 50%, then assign parameter values across a tenable range on both sides of the center. Test specificity was set at 99.5% in line with recently published analyses. There appears to be broad consensus that, for screening in an average-risk population, the true-negative rate for MCED testing should be set very high to mitigate the risk of a costly, potentially invasive, and anxiety-inducing diagnostic odyssey triggered by what turns out to be a false-positive result.
Based on these specified values for cancer point prevalence rates and MCED test characteristics, one can compute the corresponding values for PPVx, E for Cancers 1-3 and also then the FPR; the false-negative rates, denoted here as P x, E | Test- , as well as the NPV of the test; and finally P(Test+). All of these point estimates, which are displayed in the probability trees in Figure 4, enter the base-case decision model shown in Figure 5.
If the decision is to undergo MCED testing and if there is a false-negative result, corresponding estimates are required for the probability that the cancer will be symptom-detected and by stage, either upstream or downstream after testing at time t1 (see Fig. 3).
Currently, we are unaware of focused clinical evidence to guide the selection of such probability parameters, so we have proceeded with the following choices, which assume notable diversity in the patterns of symptom-driven detection.
- Cancer 1: 0.38, 0.48, and 0.14;
- Cancer 2: 0.36, 0.36, and 0.28; and
- Cancer 3: 0.22, 0.36, and 0.56.
- Cancer 1, 0.44 and 0.56;
- Cancer 2, 0.50 and 0.50; and
- Cancer 3, 0.50 and 0.50.
As noted, these probabilities are used in the calculation of QALY and cost for those decision-tree outcome nodes associated with downstream detection (see the legend for Figure 5 and Supporting Materials, Section [iv]).
If the decision is not to undergo MCED testing, estimates are similarly required for the probability of symptom detection and, by stage, during the defined upstream and downstream periods after time t1 (see Fig. 3).
- Cancer 1: 0.40, 0.50, and 0.10;
- Cancer 2: 0.40, 0.40, and 0.20; and
- Cancer 3, 0.30, 0.30, and 0.40.
- Cancer 1, 0.44 and 0.56;
- Cancer 2, 0.50 and 0.50; and
- Cancer 3, 0.50 and 0.50.
These probabilities are used in the calculation of QALY and cost for those decision-tree outcome nodes associated with downstream detection (again, see the Legend for Fig. 5 and Supporting Materials, Section [iv]).
Finally, regarding Qk and Ck, we consider each of the 3 modeled cancers in turn, indicating the assumptions and data sources that informed the derivation of QALY and cost values for the decision model. Although the resulting numbers may convey a sense of exactness, they are better viewed as representative estimates for cancers that are themselves composite representations of certain types of cancer. In that spirit, the derivation of the QALY and cost values for each of the 24 outcome nodes in Figure 5 are not detailed here; rather, the general assumptions driving these parameter selections are provided.
With pancreatic cancer as a rough prototype (for outcome calculations), data from the National Cancer Institute (NCI) Surveillance, Epidemiology, and End Results (SEER) program6 were used to inform the calculation that the health outcome advantages of being diagnosed with Cancer 1 at Earlier Stage rather than Later Stage at age 50 are real but modest, with life expectancy set at 3.10 years for detection at the Earlier Stage and at 2.58 years for detection at the Later Stage. Detailed analyses published by Tramontano et al7 and Mariotto et al8 guided our estimates of lifetime direct medical costs and QALYs for being diagnosed with Cancer 1 at Earlier and Later Stages, and the QALY calculations were informed by analyses and discussion in Tramontano et al.
With uterine cancer as a rough prototype, data and analyses from Havrilesky et al,9 the NCI SEER program,6 and the US Bureau of the Census10 were used to calculate that the life expectancy of those detected at the Earlier Stage is 31.4 years; and, for those detected at the Later Stage, it is 4.1 years. The corresponding conversion to QALYs was based on data reported by Hanmer et al11 and our additional assumptions. Estimates of lifetime direct medical cost, conditional on Earlier Stage versus Later Stage, were informed by the analyses of Havrilesky et al9 and Mariotto et al.8
With lung cancer as a rough prototype for evaluating outcomes, data were drawn from Black et al,12 the SEER program,6 and the Census Bureau10 to set the life expectancy for detection at the Earlier Stage at 9.8 years and, at the later stage, to 2.8 years. The conversion to QALYs was based in part on estimates in Black et al. Lifetime direct medical cost, conditional on stage, were informed by Black et al and also Mariotto et al.8
(We again underscore that Cancers 1, 2, and 3 are not literal representations of pancreatic, uterine, and lung cancer. Rather, to construct 3 distinctly different hypothetical cancers for which the parameter values would have some grounding in clinical reality, we arbitrarily chose those 3; other cancers could well have been selected.)
For the most likely outcome of all—namely, there is no undetected cancer at time t1—the corresponding life expectancy remaining at age 50 was set at 32 years,10 lifetime medical cost was derived from estimates for the general (noncancer) population,13 and QALY adjustments to life expectancy were based on data from Hanmer et al.11
Finally, the calculations assumed that the economic cost of the MCED test is $200; that a PET-CT scan will be done for every positive MCED test at a total cost of $2000 per scan; and also that a diagnostic workup of the cancer will be performed after a positive blood test and a positive PET-CT scan at a cost of $500.
For all 3 modeled cancers, QALY and cost estimates were discounted to present value at a real riskless annual rate of 3%, as recommended by both the original US Panel on Cost-Effectiveness in Health and Medicine and by the Second Panel.4 Regarding the decision maker's willingness-to-pay for a QALY, a recently noted14 threshold value for the US is $100,000, which was incorporated as the benchmark level for all analyses.
Cost-Effectiveness of MCED Testing: Main Findings
With all assumptions in place, Figure 5 displays the decision tree structure and parameter values that, collectively, determine the ICER associated with MCED testing within this cohort of individuals age 50 who are at risk for Cancers 1, 2 ,and 3 at t1. The bottom-line result is found in the box at the bottom of Figure 5: in these illustrative calculations, the ICER for MCED testing turns out to be $22,494 per QALY gained—well below the assumed willingness-to-pay threshold of $100,000 per QALY. This base-case result seems well in line with what Ahlquist3 anticipated when he wrote that a multiorgan approach to testing may be key to achieving cost effectiveness “by simultaneously targeting multiple tumor types and aggregating their prevalence rates” (p2).
To pursue this line of inquiry further, the cost-effectiveness calculations were re-run sequentially, assuming that, rather than there being a single MCED test for detecting Cancers 1, 2, and 3, there were instead three distinct tests for detecting Cancer 1, Cancer 2, and Cancer 3. Each cancer-specific test was assumed to have the same test characteristics as the MCED test for that cancer. Computationally, this amounted to analyzing the cost effectiveness of MCED assuming that only Cancer 1, or only Cancer 2, or only Cancer 3 was available for detection in the cohort at t1.
The resulting single-cancer ICERs are $866,015 per QALY for Cancer 1, $25,881 per QALY for Cancer 2, and $178,813 per QALY for Cancer 3. By implication, a single-cancer screening test strategy with the same basic features as the MCED strategy (vis a vis each cancer) would not be deemed cost effective for 2 of the 3 modeled cancers.
With MCED screening as presented here, the 99,025 cohort members being tested include all those with undetected Cancer 1 or Cancer 2 or Cancer 3. Hence, the total cost of screening the full cohort is being spread across a notably larger number of existing Earlier Stage cancers compared with any of the single-cancer screening scenarios considered.
- Cancer 1: Stage Shift1 = 0.000291, leading to approximately 29 additional cases of Cancer 1 detected at the Earlier Stage (relative to no MCED) of the 175 Earlier Stage cases available for detection at age t1;
- Cancer 2: Stage Shift2 = 0.000794, leading to approximately 79 additional cases of Cancer 2 detected of the 225 Earlier Stage cases available for detection; and
- Cancer 3: Stage Shift3 = 0.000316, leading to approximately 31 additional cases of Cancer 3 detected of the 125 Earlier Stage cases available for detection.
Any full-scale decision modeling analysis of the cost effectiveness of MCED testing should be accompanied by a strategically selected set of sensitivity analyses that examine the robustness of base-case conclusions to variations in key parameters. This is all the more important when the analysis is carried out parametrically, with each parameter set at a single, base-case value, as done here in Figures 4 and 5.
To illustrate the process, we carried out comparatively straightforward 1-way sensitivity analyses for 3 important parameters: the overall point prevalence of Cancers 1, 2, and 3 at time t1; test specificity; and test sensitivity for the 3 cancers. Each parameter type, in turn, was varied along the possible range of values, while all other parameters in the entire decision model were held fixed. The question in each case was: over what range of values did the computed ICER remain <$100,000, so that MCED testing remained cost effective?
The results are displayed in Figure 6. As is evident, the base-case findings about cost effectiveness are quite robust in these 1-way sensitivity analyses. The overall prevalence rate of the 3 cancers must fall below about 0.27% before the ICER exceeds $100,000; recall that the overall point prevalence rate of detectible cases of Cancers 1, 2, and 3 in the base case is 1.5%. Similarly, test specificity would need to fall below 59.2%, or test sensitivities for Cancer 1, Cancer 2, and Cancer 3 could fall jointly to 4%, 9%, and 7%, respectively, before MCED testing failed to achieve the established cost-effectiveness threshold. The complexity of such sensitivity analyses grows rapidly if and when multiple parameters are jointly varied. A state-of-the-art decision model-based analysis of MCED testing would no doubt use probabilistic sensitivity analysis, as discussed below.
Opportunities and Challenges
- What might be the impact of MCED testing on health outcomes, including life expectancy and quality-adjusted life expectancy?
- What are the implications for medical care costs?
- Under what real-world conditions would MCED testing be regarded as a cost-effective strategy?
These high-level questions motivated our illustrative analyses presented in this report. By using classical decision modeling techniques, we sought to identify the central questions, the important clinical and economic parameters, and an underlying analytical approach for evaluating the cost effectiveness of MCED testing. A central aim has been to inform the conduct—to begin to lay some foundations—for future decision model-based investigations of the cost effectiveness of MCED testing.
In our illustrative analyses, MCED testing was highly cost effective, with an ICER ($22,494 per QALY) that is well below a commonly considered threshold value for the decision maker's willingness-to-pay for a QALY ($100,000). In parallel, the analyses demonstrated how MCED testing can lead to a favorable stage shift, increasing the fraction of cases detected at an Earlier Stage for each of the hypothetical cancers here.
- The cancers included in the model would be all of those regarded as potentially detectible by MCED testing (not hypothetical composites like Cancers 1-3).
- Cancer detection would continue to be stage-specific (as in our analyses), but now stage would be defined according to criteria set by the American Joint Committee on Cancer (AJCC) or else by the NCI SEER program. Rather than a change in the probability of detection at the Earlier Stage compared with the Later Stage with MCED testing (as analyzed here), the focus would move to a change in the frequency distribution of detection across AJCC stages or SEER Summary stages. In addition, the possibility of detection at a preinvasive stage could be incorporated.
- Rather than simply 1 cohort (our individuals age 50), there would be multiple at-risk cohorts included in the model. These cohorts, perhaps better termed risk groups, would be defined (for example) by age, sex, race/ethnicity, and certain clinical variables. Key decision model parameters would be risk group–specific, including the point prevalence rates for the cancers and possibly the performance characteristics of the MCED testing strategy.
- Rather than assume a 1-time application of MCED testing across a cohort, the full-scale decision model would establish an initial age to commence screening, the time intervals for repeat MCED testing, and a stopping age. As an example, for our cohort of individuals age 50, an expanded model might assume that available and eligible cohort members would receive MCED testing every 2 years, and this would continue until age 70.
- Rather than assume that all clinical, MCED test-related, and outcome parameters in the decision model are fixed parameters known with certainty, a state-of-the-art decision model would recognize the inherent uncertainties and treat these parameters essentially as random variables, each with a mean and variance. There may also be significant nonzero covariances among certain parameters, particularly between QALYs and cost, along a given model pathway. This sets the stage for framing the conclusions about cost effectiveness in terms of the results from a probabilistic sensitivity analysis. Therefore, instead of declaring that MCED testing is cost effective or not (an up-or-down vote) based on model parameters set at their base-case values, the probabilistic approach would lead to the following type of conclusion: If the decision maker's willingness to pay for a QALY threshold value is λ, the probability that MCED testing is cost effective is P(ICER < λ), where ICER now is a random variable whose distribution is the result of the interplay of the myriad parameter distributions in the decision model.
Performing the cost-effectiveness analyses under these more realistic, and complex, assumptions will almost certainly require application of simulation modeling. Doing so will allow for much more precise modeling about when a cancer is symptom-detected, rather than the crude upstream-downstream delineation we used for computational simplicity. In addition, simulation modeling opens the way to addressing several important downstream considerations not illustrated in our calculations. These include the likelihood of cancer recurrence, as a function of stage at diagnosis for the cancer initially, and the possibility of additional primary cancers over time (plus a more explicit consideration of the influence of noncancer competing risks on outcomes). To the extent that first primary and additional downstream cancers are potentially MCED-detectible, the analyses will become ever more interesting—and complex.
Key Considerations to Accelerate Progress
As real-world investigations continue regarding the central features of MCED strategies, including (of course) test accuracy, additional efforts should proceed in parallel to evaluate the health and economic consequences of screening. These evaluations should adopt state-of-the-art modeling approaches, use the best available data, and proceed ideally within multiple at-risk population groups and global communities. In this regard, several considerations are highlighted below:
Building the Capacity for Full-Scale Decision Modeling Analysis of MCED Testing
Traditionally, single-organ cancer screening has been the primary focus of modeling analyses conducted by expert groups, such as those sponsored by the NCI’s Cancer Intervention and Surveillance Modeling Network (CISNET) program.15 Within the CISNET network, there are multiple teams of extramurally supported investigators building, testing, and curating simulation models to examine the effects of screening and treatment on outcomes currently for a total of 10 cancers: breast, bladder, cervical, colorectal, esophageal, gastric, lung, multiple myeloma, prostate, and uterine. In addition, there are published examples elsewhere of state-of-the-art, simulation-based decision modeling for endometrial cancer,9 lung cancer,12 and gastric cancer,16 among others.
What may be required for the economic evaluation of MCED testing going forward is the creation of a consortium to innovate and coordinate multiorgan modeling analyses, incorporating the panel of malignancies regarded as most likely detectible by MCED screening. This would be an unprecedented modeling effort in cancer, from both a technical and an organizational standpoint.
Complications, Limitations, and the Importance of Being Empirically Well Grounded
- Our illustrative cost-effectiveness calculations implicitly assume that every individual eventually diagnosed with Cancer 1, 2, or 3 receives the same standard of cancer care; in reality, there is considerable variation in treatment patterns within and across cancer diagnoses and population groups. Future analyses should allow for real-world variations in receipt and quality of care.
- Some decision modeling applications make an important simplifying and usually implicit assumption: If the results of the model indicate that a certain course of action (eg, get MCED testing) is optimal, the decision maker will adopt it with a probability of 1.0. In reality, uptake of any recommended procedure is variable, with factors such as out-of-pocket price, physical access to care, degree of anxiety, and anticipated discomfort influencing the individual's demand function for the procedure. For assessing the impact of MCED testing within and across population groups, projected variations in screening uptake should be incorporated into the modeling process.
- The cost-effectiveness of MCED testing will also be sensitive to variations currently in the organized cancer screening programs across global regions. The United States likely has the most extensive set of existing screening programs, with guidelines for breast, colon, cervical, and lung cancers, as well as for discussions with men of specific ages about whether to screen for prostate cancer. In other high-income countries, such as Canada and in Europe, screening for lung and prostate cancers is at an earlier stage of implementation. Organized screening programs are less common in low-income and middle-income countries, although some countries have initiated efforts to screen for breast and cervical cancers (see also the accompanying article on international perspectives). For middle-income (particularly upper middle–income) countries without existing organized screening programs, MCED testing may be attractive as a way to leap-frog the organizational and infrastructure needs of single-cancer organized screening programs. This may be the case particularly as PET/CT scanning becomes less expensive and more widely available. In low-income and lower middle–income countries where PET-CT is not widely available, MCED testing may be less cost effective, at least until the countries have sufficient capacity to treat cancers detected at earlier stages.
- There are several interesting questions about the role of MCED testing in relation to the array of existing cancer-specific detection tests. It could turn out that widespread application of an MCED test would reduce the cost effectiveness of screening for individual cancers. Yet, it is still possible for a group of tests (for single cancers as well as MCED) to have an associated overall ICER that is less than the designated cost-effectiveness threshold level. It does seem unlikely that MCED testing, as currently envisioned, would replace current tests for breast, colon, and cervical cancers. Existing tests for colon and cervical cancer can identify precancerous conditions and lead to prompt treatment, which MCED may be less able to do, whereas mammographic screening for breast cancer is more sensitive to early stage breast cancer than current versions of the MCED. A potential advantage of adding MCED as a complement to the arsenal of available tests is that it could lead to an overall increase in screening for breast, colon, and cervical cancer by, for example, being perceived as less invasive. MCED testing may also be attractive as a complement to existing lung cancer screening, which guidelines currently restrict to higher risk groups. Even if personalized screening for single cancers continues to develop, as our understanding of genetic mutations evolves, MCED testing may be a useful tool for those who do not have the known genetic markers for specific cancers. Indeed, there has been ongoing work to develop serum-based tests to detect specific cancers, including, for example, ovarian cancer in high-risk groups,17 and pancreatic,18 endometrial,9 and gastric cancers.16 Future cost-effectiveness analyses of MCED testing should bring such single-cancer blood tests into the decision model. Interesting questions then become: Does the availability of these tests reduce the cancer-specific ICERs sufficiently to render MCED itself not cost effective? Conversely, does the application of MCED lead to the conclusion that ≥1 of the single-cancer blood tests are not cost-effective? Or, might the full ensemble of all such tests, including MCED, be cost-effective?
- According to Etzioni and colleagues,19 factors that will likely influence the net clinical benefits of MCED testing if and when it is rolled out in the community include the natural history of the targeted cancers, the strategy for testing the population, and the ability to confirm a cancer signal immediately. In particular, they note that the use of PET/CT in follow-up to a positive MCED blood test not only could enhance correctly identifying the cancer tissue of origin but also could lead to incidental clinical findings. Such findings may or may not be ultimately consequential from a clinical standpoint but nonetheless could trigger additional diagnostic procedures that are costly, possibly stressful, and may yield no net clinical benefit. Those authors also note the potentially complex issues in establishing standards for how frequently individuals should undergo MCED testing, given many factors, including variations in the natural history (and thus rates of progression) of cancers that MCED may be able to detect. Moreover, for many cancers, their natural histories (which will be conditional on a range of patient factors) have not been well studied.
These real-world concerns must be acknowledged because they underscore the importance of continuing to strengthen the empirical base over time, whether through clinical trials or prospective observational studies. However, for evaluating the potential impact of MCED testing on health outcomes, cost, and cost effectiveness at any given point in time, the way forward is to bring the best data currently available to decision modeling frameworks that are well tailored for the task. As the data grow over time in quantity and quality, model performance should benefit accordingly—and, indeed, the specification of the decision models themselves may evolve. It is a dynamic process—consistent with the concept, we would say, of a learning health care system.20
Moving forward, we encourage new collaborative efforts, involving multidisciplinary teams of scientists, to build, validate, and begin to apply advanced decision modeling approaches to evaluate the impact of MCED testing on health and economic outcomes. Such analyses should draw strength from a range of scientific expertise. Key contributors would include health economists, decision scientists engaged in simulation modeling applications to cancer disease sites, biostatisticians, cancer clinical scientists, and basic-clinical translational researchers closely involved with the scientific foundations of MCED test development. Such future collaborative work, if well designed and convincingly proposed, could well merit both public and private financial support.
The aim is not simply to arrive at some up-or-down conclusion about whether MCED testing is cost effective. The more important focus should be on identifying the set (or range) of clinical and economic assumptions under which such screening meets current benchmark criteria for being cost effective. An essential feature of a decision modeling approach is that it allows one to investigate the robustness of conclusions about cost effectiveness to variations in those critical assumptions. The time is ripe to embark upon such analyses in a comprehensive, rapid, and coordinated way.
No specific funding was disclosed.
Conflict of Interest Disclosures
Joseph Lipscomb reports travel support from the International Union Against Cancer, the American Cancer Society, and the Mayo Clinic during the conduct of the study and grants from the National Cancer Institute and the Centers for Disease Control and Prevention outside the submitted work. Susan Horton reports travel support from the International Union Against Cancer, the American Cancer Society, and the Mayo Clinic during the conduct of the study; grants from the Canadian Institutes for Health Research, the Institute for Catastrophic Loss Reduction, and the Brocher Foundation; and personal fees from PATH outside the submitted work. Albert Kuo reports salary support from the John Templeton Foundation during the conduct of the study and has a patent pending through Johns Hopkins University outside the submitted work. Cristian Tomasetti reports grants from the John Templeton Foundation during the conduct of the study and royalties under an agreement between Exact Sciences Corporation and Johns Hopkins University outside the submitted work.
|cncr33980-sup-0001-Supinfo.docxapplication/docx, 20.4 KB||Supplementary Material|
Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.
- 1, , , et al. Feasibility of blood testing combined with PET-CT to screen and guide intervention. Science. 2020; 369:eabb9601.
- 2, , , , , CCGA Consortium. Sensitive and specific multi-cancer detection and localization using methylation signatures in cell-free DNA. Ann Oncol. 2020; 31: 745-759.
- 3. Universal cancer screening: revolutionary, rational, and realizable. NPJ Precis Oncol. 2018; 2: 23. doi:10.1038/s41698-018-0066-x
- 4, , , et al. Recommendations for conduct, methodological practices, and reporting of cost-effectiveness analyses: Second Panel on Cost-Effectiveness in Health and Medicine. JAMA. 2016; 316: 1093-1103.
- 5 Surveillance Research Program, Division of Cancer Control and Population Sciences, National Cancer Institute. Cancer Query System: Cancer Prevalence Database. Accessed November 1, 2020. https://surveillance.cancer.gov/prevalence/canques.html
- 6 Surveillance Research Program, Division of Cancer Control and Population Sciences, National Cancer Institute. Cancer Query System: SEER Survival Statistics. Accessed November 1, 2020. https://seer.cancer.gov/canques/survival.html
- 7, , , et al. Pancreatic cancer treatment costs, including patient liability, by phase of care and treatment modality, 2000-2013. Medicine (Baltimore). 2019; 98:e18082.
- 8, , , , . Projections of the cost of cancer care in the United States: 2010-2020. J Natl Cancer Inst. 2011; 103: 117-128.
- 9, , . Cost-effectiveness analysis of annual screening strategies for endometrial cancer. Am J Obstet Gynecol. 2009; 200: 640.e1-640.e8.
- 10, , . Appendix Table 1: Life Table for the Total Population in the United States: 2017. In: Living Longer: Historical and Projected Life Expectancy in the United States, 1960 to 2060. Current Population Report P25-1145. US Census Bureau; 2020: 16-17.
- 11, , , , . Report of nationally representative values for the noninstitutionalized U.S. adult population for 7 health-related quality-of-life scores. Med Decis Making. 2006; 26: 391-400.
- 12, , , et al. Cost-effectiveness of CT screening in the National Lung Screening Trial. N Engl J Med. 2014; 371: 1793-1802.
- 13, . Peterson-KFF Health System Tracker: How Do Health Expenditures Vary Across the Population? Peterson-Kaiser Family Foundation; 2019. Accessed November 15, 2020. https://www.healthsystemtracker.org/chart-collection/health-expenditures-vary-across-population/#item-start
- 14, , . Updating cost-effectiveness—the curious resilience of the $50,000-per-QALY threshold. N Engl J Med. 2014; 371: 796-797.
- 15 National Cancer Institute. Cancer Intervention and Surveillance Modeling Network (CISNET). Accessed March 2, 2021. https://cisnet.cancer.gov
- 16, , , , , . Evaluating the use of microRNA blood tests for gastric cancer screening in a stratified population-level screening program: an early model-based cost-effectiveness analysis. Value Health. 2020; 23: 1171-1179.
- 17, , , et al. Development of a longitudinal two-biomarker algorithm for early detection of ovarian cancer in women with BRCA mutations. Gynecol Oncol. 2020; 159: 804-810.
- 18, , , et al. Combined circulating tumor DNA and protein biomarker-based biopsy for the earlier detection of pancreatic cancers. Proc Natl Acad Sci U S A. 2017; 114: 10202-10207.
- 19, , . Multi-cancer early detection: learning from the past to meet the future. J Natl Cancer Inst. Published online August 27, 2021. doi:10.1093/jnci/djab168
- 20 Institute of Medicine. A Foundation for Evidence-Drive Practice: A Rapid Learning System for Cancer Care: Workshop Summary. The National Academies Press; 2010.