Clinical and Economic Impact of Tailoring Screening to Predicted Colorectal Cancer Risk: A Decision Analytic Modeling Study

This article is featured in Highlights of This Issue, p. 265

Personalized preventive medicine may optimize clinical outcomes and population-wide health, while guiding judicious use of limited and costly resources, compared with a preventive approach that is applied uniformly. Colorectal cancer screening is an ideal case study for the potential of personalized preventive medicine. Screening decreases colorectal cancer incidence and mortality (1–7). Multiple organizations endorse colorectal cancer screening for the “average-risk” population (8–10). Higher intensity screening is recommended for persons identified as high risk based on family history (9, 11–13). Age is used to define a younger group in which colorectal cancer risk is low enough that screening is not yet recommended (8–10). The American Cancer Society's recent recommendation that adults at average risk of colorectal cancer undergo regular screening beginning at age 45 instead of 50 (10) has spurred debate about the optimal use of screening resources, including the potential role of risk stratification to guide colorectal cancer screening.

The “average risk” population includes people across a spectrum of colorectal cancer risk, but methods to determine an individual's risk remain to be perfected. Because most people will never develop or die from colorectal cancer, the vast majority of colorectal cancer screening participants derive no benefit, but are exposed to risks and incur costs. Tailoring screening to predicted colorectal cancer risk could improve the clinical effectiveness of colorectal cancer screening while minimizing complications and optimizing resource use (14). Various colorectal cancer risk prediction tools with modest discriminatory ability exist (14–16), but they are not used widely.

We used a validated decision analytic model of colorectal cancer screening to explore the potential impact of colorectal cancer screening tailored to the level of predicted lifetime colorectal cancer risk compared with the current usual care of uniform screening intensity for the “average risk” population. Our concept of a risk prediction tool is broad: it can be based on clinical criteria, single or multiple genetic or other biomarkers, or it can be a hybrid (14–16). We first determined colorectal cancer risk levels that could be defined as “low risk” or “high risk” based on threshold analyses. We then compared risk stratification and tailored screening intensity versus usual care in the context of a colonoscopy-based or a fecal immunochemical test (FIT)-based screening program, consistent with the recommendation of the U.S. Multisociety Task Force on Colorectal Cancer to follow a risk-stratified approach regarding test choice, with colonoscopy offered to higher risk and FIT to lower risk persons (9). We explored the influence of a risk-prediction tool's discriminatory ability and cost, as well as the assumed distribution of true colorectal cancer risk represented in the “average risk” population.

We adapted a validated decision analytic Markov cohort model of colorectal cancer screening in the general U.S. population (constructed in TreeAge Pro, TreeAge Software Inc.; refs. 17–19; Supplementary Fig. S1) to incorporate a spectrum of colorectal cancer risk, risk stratification, and tailored screening. Because the distribution of true or actual colorectal cancer risk within the “average risk” population is not known, we modeled six plausible distributions (refs. 20, 21; generated in SAS, SAS Institute Inc.; Fig. 1; Table 1; Supplementary Data). Under each scenario, individuals were assigned to a tailored colorectal cancer screening strategy from ages 50 to 80 years based on their predicted colorectal cancer risk score. While an individual's actual risk governed the model-simulated colorectal cancer risk over time, the predicted risk assigned by a prediction tool governed the recommended screening intensity.

The American Cancer Society's recent recommendation (10) implies that tailored screening could occur before age 50, but for this study we focused on tailored screening intensity instead of starting age because currently many persons remain unscreened even after age 50, the screening initiation age is only beginning to be reexamined, ascertainment of family history, which is the best recognized risk factor, is very poor in current clinical practice (22), and most existing risk-prediction models were not developed to address younger ages (15, 16). We have recently studied the potential clinical impact and cost-effectiveness of initiating colorectal cancer screening at age 45 in a separate study (23).

We recognize that actual colorectal cancer risk for an individual may change over time, and that predicted colorectal cancer risk could be informed by the results of past screening. Because reliable data are not currently available to inform such nuanced modeling, we made the simplifying assumption that risk-stratification was performed only at simulation entry, akin to baseline ascertainment of risk for persons with genetic cancer predisposition syndromes. However, surveillance in our model is informed by previous findings. Regardless of the initial predicted risk, persons who have a large polyp undergo surveillance at 3 years, and those with a small polyp at 5 years, consistent with current guidelines. On the other hand, persons with higher predicted risk do not transition to less intense testing after normal colonoscopy, consistent with current approaches to patients with family history of colorectal cancer or hereditary syndromes.

We conceptualized tailored screening as the stratification of the “average risk” population into a “low risk” tier that could be screened at a lower intensity, a “moderate risk” tier that could be screened as recommended for the “average risk” population, and a “high risk” tier that could be screened at a higher intensity. Although highly tailored personalized recommendations across a spectrum of screening tests and test intervals could be contemplated, we deliberately modeled a simpler scheme that is feasible to implement in practice.

There is no universal agreement on the levels of colorectal cancer risk that justify lower intensity screening, or that warrant higher intensity screening. Therefore, we first performed threshold analysis to identify levels of colorectal cancer risk at which screening of different intensity could be supported, on the basis of a conservative willingness-to-pay threshold of $50,000/quality-adjusted life-year (QALY) gained (24).

We first considered the most optimistic case, in which a risk-prediction tool could generate a predicted colorectal cancer risk without misclassifying individuals (i.e., predicted risk always matched actual risk) at a negligible cost (e.g., applying a computerized algorithm during an existing clinical encounter). We then explored the consequences of using an imperfect tool with an inherent misclassification rate (14–16), at an incremental cost (e.g., performing laboratory assays).

Clinical and economic outcomes were estimated in the model in TreeAge Pro. These estimates were then aggregated in analyses performed in SAS and in Excel (Microsoft Corporation).

The original model, data sources, inputs, and validation against the Minnesota Colon Cancer Control Study (25, 26), United Kingdom Flexible Sigmoidoscopy Trial (27), SCORE Trial (28), and PLCO Cancer Screening Trial (29) have been detailed previously (refs. 17, 18, 30; Table 1; Supplementary Fig. S1).

We generated six hypothetical populations with different distributions of actual colorectal cancer risk (20, 21) (four unimodal beta distributions and two bimodal distributions made up of two overlapping beta distributions) and an average lifetime risk of colorectal cancer of 5.9% from age 50, based on Surveillance Epidemiology and End Results data from 1992 to 1994, before widespread colorectal cancer screening (ref. 31; Fig. 1; Table 2; Supplementary Table S1). Each distribution is characterized by different fractions of the population contained in the low, moderate, and high risk tiers.

For the base case, we constructed a population with a unimodal beta distribution, in which the upper 2.5%-ile end of the population distribution was 2.5-fold the average risk, based on the expected range of risk as informed by the impact of family history of colorectal cancer and previous work on breast cancer (ref. 20; “Unimodal 2.5x”; Fig. 1; Table 2; Supplementary Data). Five other plausible distributions were explored in sensitivity analyses, three unimodal and two bimodal (refs. 20, 21; Fig. 1; Table 2; Supplementary Data). These were generated based on 1,000,000 persons (Supplementary Data). For the unimodal distributions, progressively more individuals fell into the lower risk and higher risk tiers as the actual colorectal cancer risk distribution became broader (Fig. 1; Table 2). For the bimodal distributions, a high risk subgroup was modeled as a discrete subpopulation, as opposed to high risk being reflected only by the upper tail of a unimodal risk distribution (Fig. 1).

We validated that each distribution can underlie the observed age-dependent incidence of colorectal cancer without screening (ref. 31; Supplementary Fig. S2).

We compared tailored screening to two established screening strategies. We compared “tailored colonoscopy” consisting of once-only colonoscopy at age 60 in persons categorized as low risk, colonoscopy every 10 years in persons categorized as moderate risk, and colonoscopy every 5 years in persons categorized as high risk versus the usual care of “colonoscopy every 10 years.” We compared “tailored FIT/colonoscopy” consisting of annual FIT in persons categorized as low or moderate risk, and colonoscopy every 5 years in persons categorized as high risk versus the usual care of “annual FIT.”

We selected these risk-specific–tailored strategies based on the low risk of metachronous neoplasia in persons with normal colonoscopy (32), guidelines recommending colonoscopy every 5 years in persons with a family history of colorectal cancer (9), clinical judgment, and the predicted ease of implementation in clinical practice.

Most colorectal cancer risk prediction tools have not been validated prospectively. An exception is the NCI Colorectal Cancer Risk Assessment Tool. Two studies have shown that it can stratify individuals into quintiles of advanced neoplasia prevalence at colonoscopy ranging from 2% to 6% through 17% to 18% in the lowest and highest quintiles, respectively (33, 34). However, the discriminatory accuracy of the tool is only modest (33, 35), with substantial overlap between the risk score distributions of persons with or without advanced neoplasia (33). In our previous study, the fraction of persons with advanced neoplasia who were assigned by the 10-year risk score to the lowest through highest quintile of risk were 10%, 14%, 21%, 26%, and 29%, respectively (33). This illustrates that: (i) misclassification was substantial (24% of persons with advanced neoplasia had a predicted risk in the two lowest quintiles, and 45% in the lowest three quintiles), and (ii) misclassification can be extreme (some persons with advanced neoplasia were inappropriately placed in the lowest quintile). Similarly, in the study by Imperiale and colleagues, the fraction of persons with advanced neoplasia who had risk scores below the median was 23%–29% for the 5-year to 20-year risk scores (34).

Consistent with these observations, we explored misclassification rates of 10%–40% for all hypothetical risk distributions. In our primary approach, for any given misclassification rate, we assumed that for each risk tier, one half of the misclassified individuals would be assigned inappropriately to each one of the other two risk tiers (illustrated in Supplementary Fig. S3). For instance, misclassification led to a given fraction of individuals at actual high colorectal cancer risk receiving colonoscopy only once at age 60—the intensity tailored to low risk persons.

We analyzed the Unimodal 2.5x distribution with 20% misclassification as a case example to contrast the impact of overestimation versus underestimation of risk (Supplementary Fig. S4). Finally, we also explored a secondary scenario with less extreme 20% misclassification in this distribution. In this case, only the one half of a tier's risk scores that is adjacent to a cut-off point between tiers was subject to misclassification; and misclassification was possible only to the adjacent tier (Supplementary Fig. S5).

Base case cost inputs were derived from Medicare reimbursement rates (http://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/ClinicalLabFeeSched/, http://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/AcuteInpatientPPS/, and http://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/PhysicianFeeSched/) and estimated colorectal cancer care costs (36) accounting for colonoscopy site of service (36), and updated to 2013 dollars using the medical component of the consumer price index (Table 1).

The principal model outputs were QALYs and costs per person with a lifetime horizon (37). Future QALYs and costs were discounted by 3% annually (38). In calculating QALYs, health state utilities for colorectal cancer by stage (Table 1) were applied for 5 years after colorectal cancer diagnosis. We estimated colorectal cancer cases by stage and colorectal cancer–related deaths, and lifetime number of colonoscopies per person in cohorts of 100,000 persons.

Analyses from the health sector perspective of a third-party payer were performed. Incremental cost-effectiveness ratios were calculated (37). We used the commonly accepted willingness-to-pay threshold of $50,000–100,000/QALY gained (24).

Our previous extensive sensitivity analyses on all model inputs have demonstrated that varying these inputs over plausible ranges does not affect the major conclusions regarding FIT or colonoscopy (17–19). Here, we performed detailed sensitivity analyses on the variables judged a priori to be critical to this research question: the distribution of actual colorectal cancer risk in the average risk population, the risk-prediction tool's discriminatory ability, and its cost. The incremental cost of the risk-prediction tool could reflect tool application costs during a clinic visit or during outreach, fees for laboratory-based testing, or both. Tool cost was assigned at simulation entry.

At willingness-to-pay threshold of $50,000/QALY gained, the cutoff for low risk emerged as ≤3% actual lifetime colorectal cancer risk from age 50. For these persons, colonoscopy every 10 years was costly with respect to its benefits versus once-only colonoscopy at age 60 (Supplementary Table S2). Similarly, the cutoff for high risk emerged as ≥12% actual lifetime colorectal cancer risk from age 50. For these persons, colonoscopy every 5 years was cost-effective versus colonoscopy every 10 years (Supplementary Table S2).

Coincidentally, these thresholds corresponded to approximately half and approximately twice the population average colorectal cancer risk, respectively. Current clinical guidelines recommend colonoscopy every 5 years for persons with a first-degree relative with colorectal cancer, which confers approximately twice the average risk (9). This clinical standard of care provided external validation to our definition of the high risk tier.

Tailored screening was preferred over usual care when there was no misclassification of individuals and no cost to apply the risk-prediction tool. The clinical and economic rationale supporting tailored screening differed between the two tailored screening strategies. Tailored colonoscopy prevented slightly fewer colorectal cancer cases and deaths, and yielded slightly shorter quality-adjusted life expectancy, than colonoscopy every 10 years, but the relatively small clinical advantages of colonoscopy every 10 years were achieved at the relatively high cost of $146,800/QALY gained, coupled with a greater demand for colonoscopy (Table 3). In contrast, tailored FIT/colonoscopy prevented more colorectal cancer cases and deaths, and yielded longer quality-adjusted life expectancy, than annual FIT, and these advantages were achieved at a relatively cost-effective $48,400/QALY gained despite a higher demand for colonoscopy (Table 3).

The precise balance between the incremental clinical benefits and costs of tailored screening compared with usual care depended on the assumed underlying actual colorectal cancer risk distribution in the average risk population. However, the conclusions were robust when there was no misclassification of individuals and no cost to apply the risk-prediction tool. Under five of the six risk distributions, tailored colonoscopy was slightly less effective than colonoscopy every 10 years, but colonoscopy every 10 years was relative costly ($90,200–$889,000/QALY gained); only under the bimodal 10/80/10 risk distribution, tailored colonoscopy was both more effective and less costly than colonoscopy every 10 years (Table 4; Supplementary Fig. S6). Under all six risk distributions, tailored FIT/colonoscopy was slightly more effective and cost-effective compared with annual FIT ($10,600–$60,000/QALY gained; Table 4; Supplementary Fig. S6).

Even relatively modest levels of colorectal cancer risk misclassification with the risk-prediction tool resulted in uniform screening being preferred over tailored screening under all six colorectal cancer risk distributions. As the colorectal cancer risk misclassification rate increased from 10% to 40%, colonoscopy every 10 years became progressively more effective and cost-effective compared with tailored colonoscopy (Fig. 2A; Supplementary Tables S3 and S4), and tailored FIT/colonoscopy became progressively less effective, under most distributions, and less cost-effective compared with annual FIT (Fig. 2B; Supplementary Tables S3 and S4). Specific thresholds can be gleaned from Fig. 2.

In the Unimodal 2.5x distribution with 20% misclassification as a case example, overestimation (Supplementary Fig. S4A) and underestimation (Supplementary Fig. S4B) of risk accounted for an absolute 12% and 8% misclassification rate, respectively. Underestimation drove the decrease in tailored colonoscopy's effectiveness (0.0033 QALY/person lost vs. 0.0009 QALY/person gained with overestimation), and overestimation drove the increase in cost (incremental $138/person incurred vs. $35/person saved with underestimation). For tailored FIT/colonoscopy, overestimation and underestimation affected effectiveness minimally (0.0003 QALY/person gained and 0.0002 QALY/person lost, respectively), and overestimation drove the increase in cost (incremental $161/person incurred vs. $12/person saved with underestimation).

In the secondary scenario, with less extreme 20% misclassification in the Unimodal 2.5x distribution (Supplementary Fig. S5), colonoscopy every 10 years was still preferred versus tailored colonoscopy ($62,000 vs. $32,000/QALY gained with the original more extreme 20% misclassification). In contrast, tailored FIT/colonoscopy emerged as preferred versus annual FIT ($71,900 vs. $158,000/QALY gained with the original more extreme 20% misclassification).

As the risk-prediction tool cost increased, colonoscopy every 10 years became progressively more cost-effective compared with tailored colonoscopy, and tailored FIT/colonoscopy became progressively more costly compared with annual FIT (Supplementary Table S5). Assuming the Bimodal 10/80/10 risk distribution, the only distribution under which tailored colonoscopy was more effective than colonoscopy every 10 years in the base case, and assuming no risk misclassification, the clinical advantages of tailored colonoscopy were realized at incremental costs of $70,000–$195,000/QALY gained when the risk-prediction tool cost was $100–$125.

The combination of relatively modest risk misclassification rates and risk-prediction tool costs tended to make uniform screening the preferred approach over tailored screening (Fig. 2C and D).

The results of this decision analysis clarify the requisites for colorectal cancer risk-prediction and tailored screening to be beneficial at the population level, and they highlight principles that may be generalizable to personalized preventive medicine in cancer control more broadly.

Under various plausible assumed actual colorectal cancer risk distributions in the “average risk” population, tailored screening was preferred over the usual care of uniform screening—assuming no misclassification of risk and no cost to apply the risk-prediction tool. A similar theme was explored by a recent study focusing on family history of colorectal cancer, which suggested that tailoring can be guided by the details of a person's family history of colorectal cancer (39). In this study, however, even relatively modest colorectal cancer risk misclassification rates in the “average risk” population or relatively modest risk-prediction tool costs led to uniform screening emerging as the preferred approach.

The results illustrate the considerations that could make tailored colorectal cancer screening a preferred approach for the “average risk” population in routine clinical practice. Tailored colonoscopy was preferred over colonoscopy every 10 years, assuming no misclassification or risk-prediction tool cost, because it was only slightly less effective but substantially less costly under most risk distributions. In contrast, tailored FIT/colonoscopy was preferred over annual FIT because it was slightly more effective at an acceptable incremental cost. Preventive interventions that achieve both improved outcomes and lower costs are ideal, but rare (40).

These results suggest that the currently available colorectal cancer risk-prediction tools, which have an area under the receiver operating characteristic curve (AUROC) of 0.6–0.78 (16), may lack the necessary level of accuracy to justify their application to select some persons into lower-than-average intensity screening, and others into higher-than-average intensity screening (21, 41). The NCI Colorectal Cancer Risk Assessment Tool, for instance, can stratify individuals into quintiles of current advanced neoplasia (33, 34), but its discriminatory accuracy is only modest, with an AUROC of 0.61 [95% confidence interval (CI), 0.60–0.62] for men and 0.61 (95% CI, 0.59–0.62) for women (33, 35). Relating this to the misclassification rates that we modeled, the fraction of persons with advanced neoplasia who were assigned by the 10-year risk score to the lowest through highest quintile of risk in our previous study were 10%, 14%, 21%, 26%, and 29%, respectively (33). Similarly, in another study, the fractions of persons with advanced neoplasia who had risk scores below the median were 23%–29% (34). Thus, with this tool, misclassification into lower risk tiers occurs at substantial rates among persons with advanced neoplasia, who most clinicians would want to see in higher intensity screening.

In our primary analysis of misclassification, assuming the risk prediction tool is free, colonoscopy every 10 years was preferred over tailored colonoscopy under all six hypothetical risk distributions when the misclassification rate was under approximately 20% at a willingness-to-pay threshold of $100,000/QALY (Fig. 2A). Tailored FIT/colonoscopy was preferred over annual FIT under all six hypothetical risk distributions when the misclassification rate was under approximately 5%, and under five of the six distributions when the misclassification rate was under approximately 10%, at a willingness-to-pay threshold of $100,000/QALY gained (Fig. 2B).

The principal determinants of the consequences of misclassification were increased cost attributable to overestimation of risk and decreased effectiveness attributable to underestimation of risk. This suggests that when applying an imperfect risk score, more intense screening in persons with higher risk scores will probably improve cancer prevention efforts, in exchange for higher costs and an increase in complications, which are rare overall. However, recommending less intense screening in persons misclassified as lower risk could adversely affect cancer prevention (14).

It remains to be clarified how much of the colorectal cancer burden in the population is predictable (42), and whether prediction tools can achieve the necessary discriminatory accuracy (21, 41). Risk-prediction algorithms that do not require specialized assays may approach a negligible cost, but it is not clear that demographic and clinical attributes alone can achieve the necessary levels of risk-prediction accuracy (14). The use of laboratory assays, alone or in combination with clinical factors, could improve risk-prediction accuracy, but their application will entail population-wide expenditures to achieve risk stratification.

Public policy on cancer control must consider the challenges of implementation as much as the accuracy of tests and the effectiveness of interventions. We have shown previously that even small differences in screening participation can overwhelm differences in screening test performance characteristics (19, 30). If risk-stratification and tailored screening prove too onerous for patients, clinicians, and health systems, so that overall screening participation rates suffer, net harm could result. Not only must a colorectal cancer risk-stratification tool be highly accurate and relatively inexpensive, but it must also be implemented without inadvertently decreasing screening participation.

This analysis has limitations. First, we modeled colorectal cancer assessment at simulation entry, without explicitly modeling dynamic reassessment of colorectal cancer risk over time with a risk-prediction tool. Nonetheless, our simulations also reflect a scenario in which colorectal cancer risk is actually reassessed dynamically and individuals shift into different risk tiers over time, but with the overall population proportions in each risk tier remaining the same as in our simplified scenario. Our model does account for changes in management based on polyp history, thus reflecting this component of dynamic risk reassessment for surveillance. Second, we did not consider risk prediction starting at ages younger than 50, which could lead to earlier initiation of screening in higher risk persons, for the reasons explained in the Materials and Methods section. We refer readers interested in the question of colorectal cancer screening at age 45 to our recent analysis (23). Third, we did not model different approaches based on gender alone, reasoning that any independent effect of gender on colorectal cancer risk would be incorporated into the risk prediction (15, 16). Fourth, we did not model different participation levels in risk stratification, screening or surveillance, or the impact of risk stratification on subsequent screening behavior. The incremental clinical and economic differences between strategies would be lower with uniformly lower participation levels, and it is conceivable that the preferred strategy could change if participation levels differed markedly between strategies, for instance, if risk assessment encouraged or somehow discouraged overall screening participation. Finally, we did not consider foregoing screening in persons at very low predicted risk.

It is said that the perfect is the enemy of the good. In colorectal cancer screening, imperfect personalization could be the enemy of a simpler, uniform population-wide approach. The pursuit of personalized colorectal cancer prevention is a worthy goal that is motivated by the promise of optimizing population health and resource utilization. This pursuit must be informed by a sober appreciation of the requisites (43), and of the implementation challenges. Our results suggest that colorectal cancer risk stratification to intensify or relax screening recommendations in subgroups within the current “average risk” population must be highly accurate and relatively inexpensive to realize greater population-wide benefits than those achieved with uniform screening recommendations.

U. Ladabaum is a consultant for Medtronic, Motus, and Clinical Genomics and has ownership interest in Universal Dx. No potential conflicts of interest were disclosed by the other authors.

The study was funded by a Stanford Cancer Institute Innovation Award, which had no role in the identification, design, conduct, or reporting of the analysis.

Conception and design: U. Ladabaum, A. Mitani

Development of methodology: U. Ladabaum, M. Desai

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): U. Ladabaum, A. Mannalithara

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): U. Ladabaum, A. Mannalithara, A. Mitani, M. Desai

Writing, review, and/or revision of the manuscript: U. Ladabaum, A. Mannalithara, A. Mitani, M. Desai

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): U. Ladabaum, A. Mannalithara

Study supervision: U. Ladabaum

The study was funded by a Stanford Cancer Institute Innovation Award. The authors thank Dr. Alice Whittemore for her recommendations, review, and suggestions.

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.