Merkel cell carcinoma recurrence risk estimation is improved by integrating factors beyond cancer stage: a multivariable model and web-based calculator

Abstract

Background

Merkel cell carcinoma (MCC) recurs in 40% of patients. In addition to stage, factors known to affect recurrence risk include: sex, immunosuppression, unknown primary status, age, site of primary tumor, and time since diagnosis.

Purpose

Create a multivariable model and web-based calculator to predict MCC recurrence risk more accurately than stage alone.

Methods

Data from 618 patients in a prospective cohort were used in a competing risk regression model to estimate recurrence risk using stage and other factors.

Results

In this multivariable model, the most impactful recurrence risk factors were: AJCC stage (p<0.001), immunosuppression (hazard ratio 2.05; p<0.001), male sex (1.59; p=0.003) and unknown primary (0.65; p=0.064). Compared to stage alone, the model improved prognostic accuracy (concordance index for two-year risk, 0.66 vs. 0.70; p<0.001), and modified estimated recurrence risk by up to 4-fold (18% for low-risk stage IIIA vs. 78% for high-risk IIIA over five years).

Limitations

Lack of an external data set for model validation.

Conclusion / Relevance

As demonstrated by this multivariable model, accurate recurrence risk prediction requires integration of factors beyond stage. An online calculator based on this model (at merkelcell.org/recur) integrates time since diagnosis and provides new data for optimizing surveillance for MCC patients.

Capsule Summary

• We describe a personalized recurrence risk calculator for Merkel cell carcinoma that integrates stage, sex, age, primary tumor site, immunosuppression, and unknown primary status to improve prognostic accuracy.

• A web-based tool also integrates time since initial treatment, providing new guidance for optimized surveillance of Merkel cell carcinoma.

Introduction

Merkel cell carcinoma (MCC) is a relatively rare but aggressive skin cancer. MCC incidence in the US is increasing with 3,284 new cases predicted in 2025.^1–3 MCC has a high risk of recurrence (40%), with >90% of recurrences arising within the first 3 years after diagnosis.⁴ American Joint Committee on Cancer (AJCC) stage at diagnosis provides important prognostic information, and higher stage confers a higher recurrence risk (at 5 years, 72% of patients with stage IV experienced recurrence vs. 20% of those with stage I).⁴ However, multiple patient and tumor characteristics not included in stage are considered poor prognostic factors: male sex, advanced age, immune suppression, and head/neck primary tumor location. Although stage-specific recurrence data has recently been published⁴, the combined prognostic impact of stage, patient and tumor characteristics, and time since diagnosis has not been described. Adding these factors significantly alters risk as calculated by stage alone, and thus can better guide surveillance duration and intensity. Data-driven surveillance can help prevent low-risk patients from undergoing unnecessary imaging studies and examinations as well as prevent high-risk patients from experiencing disease progression that could have otherwise been detected and treated sooner.

Here we present a multivariable model which includes stage and other risk factors for estimating MCC recurrence risk, and an online calculator that explicitly integrates time since diagnosis, an important determinant of residual recurrence risk. While the model is derived from a patient cohort that was described together with the univariable impact of each factor⁴, new multivariable analyses were required to generate and evaluate a personalized risk model. The unique recurrence risk estimates generated from this calculator will help clinicians differentiate high-risk patients meriting close follow-up versus low-risk patients as well as determine when de-escalation of surveillance may be appropriate.

Methods

Patients with pathologically confirmed MCC were prospectively enrolled between January 2003 and April 2019 in an institutional review board-approved repository maintained at the University of Washington in Seattle with written informed consent provided by participants. Full inclusion and exclusion criteria are outlined in the Mendeley Supplemental Methods. Patient and tumor characteristics shown to be related to prognosis in prior MCC studies were included in this analysis: AJCC stage, sex, age, immune suppression status, unknown primary status, and primary tumor site. Patients were considered immune suppressed if, at the time of MCC diagnosis, they had CLL, HIV/AIDS, solid organ transplant, autoimmune diseases (rheumatoid arthritis, lupus, Crohn’s disease, ulcerative colitis, and Lambert-Eaton Syndrome), or other hematologic malignancy (non-Hodgkin’s lymphomas, multiple myeloma, and mycosis fungoides).⁵

The primary endpoint was time to first MCC recurrence after initial treatment. MCC recurrence was defined as reappearance of disease in imaging or clinical exam; or progression of existing disease after diagnosis and initial treatment (predominately surgery and radiation therapy). This study considered only the first recurrence observed for each patient. The date of completion of initial treatment was not systematically documented in the repository and was approximated as 90 days after the initial diagnosis of MCC. This approximation was used because the median duration of initial treatment (surgery, radiation and systemic therapy) was approximately 90 days and we sought a cohort representative of post-treatment patients. Time to recurrence was calculated from the approximate date of completion of initial treatment to the date that first recurrent disease was clinically detected. Similarly, time to death was calculated as the interval between dates of completion of initial treatment and death. Patients who did not experience a recurrence or die during their follow-up time were censored using the date of their last contact. Death from non-MCC causes was considered a competing risk throughout the analysis. If a patient died from MCC without a recurrence clinically determined first, their death date was used as the date of recurrence. Additional details are provided in Mendeley Supplementary Methods.

The overall risk of recurrence was estimated using the cumulative incidence function estimator, accounting for the competing risk of death. A multivariable model for MCC recurrence was fit using Fine & Gray competing risk regression for all collected predictor variables, with death treated as a competing risk. Another competing risk regression model was fit using AJCC stage as the only predictor so the gain in performance from adding additional patient and tumor characteristics to stage could be evaluated. No predictors had missing values. Model performance (discrimination and calibration) was evaluated using a concordance index that accounts for competing risks (discrimination)⁶, the index of prediction accuracy (IPA; a combination of discrimination and calibration performance)⁷, and calibration curves (Mendeley Supplemental Figure 2). Internal validation techniques, including bootstrap-based optimism adjustments and 10-fold cross-validation, were utilized to efficiently achieve relatively unbiased estimates of performance without an additional testing set.^7–9 Statistical calculations were performed using R (versions 3.6.1 and 4.0.3; R Foundation for Statistical Computing, Vienna, Austria). The R Shiny package was used to implement the multivariable model in a web-based calculator for public use.¹⁰ The mstate package (version 0.3.2)¹¹ and rms package (version 6.3, Harrell Jr., Franke E, 2022) were used to generate a nomogram of the multivariable model that accounts for competing risks.¹²

Results

MCC cohort

We analyzed a 618 patient cohort from our Seattle-based MCC repository that was previously described together with a univariable analysis of individual risk factors and outcomes⁴ (Mendeley Supplemental Figure 1). Over the follow-up period there were 187 deaths, of which 121 were due to MCC. Fifty-two deaths occurred prior to a recurrence (competing risk).

Multivariable model

The multivariable model for MCC recurrence risk with all predictors is summarized in Figure 1. The hazard ratios corresponding to AJCC stage ranged from 1.65 (95% CI: 0.84-3.23) for clinical stage I (reference: P-I) to 7.34 (95% CI: 3.86-14.0) for stage IV. After adjusting for stage, immunosuppression (HR: 2.05, 95% CI: 1.42-2.97), female sex (HR: 0.63, 95% CI: 0.46-0.86), and unknown primary (HR: 0.65, 95% CI: 0.41-1.03) had strong estimated associations with recurrence. In contrast, age (HR: 1.10 per 10-year increase, 95% CI: 0.96-1.25), a trunk primary tumor (HR: 0.82, 95% CI: 0.53-1.37, reference: extremity), and a head/heck primary tumor (HR: 0.94, 95% CI: 0.65-1.37, reference: extremity) had weaker associations with recurrence. Compared to previously published univariable analyses using data from this cohort⁴, each individual factor had a similar or lower impact on recurrence risk after adjusting for the other factors in the multivariable model.

Figure 1:

Using a multivariable competing risk regression model, hazard ratios for stage and other factors were calculated and represent the impact of each factor on risk of MCC recurrence. The forest plot displays each hazard ratio and associated 95% confidence interval. Ref = reference level for each categorical variable, HR = hazard ratio, CI = confidence interval. *Patients with an unknown primary typically present with disease in a lymph node and do not have any detectable skin lesion at diagnosis.

The optimism-adjusted concordance index estimates for recurrence were 0.70 (95% CI: 0.67-0.74) at 2 years and 0.68 (95% CI: 0.65-0.72) at 5 years. The optimism-adjusted IPA estimates were 0.14 at 2 years (95% CI: 0.10-0.22) and 0.11 at 5 years (95% CI: 0.07-0.20). Overall, the multivariable model appeared to be well-calibrated in our population-based calibration plots (Mendeley Supplemental Figure 2). Figure 2 summarizes the multivariable model in a nomogram to allow graphical computation of an individual patient’s recurrence risk. A web-based implementation of the multivariable model is shown in Figure 3.

Figure 2:

Nomogram for 5-year recurrence risk of Merkel cell carcinoma (MCC) after initial treatment. Locate a given patient’s sex on the sex axis. Draw a straight line up to the point axis to determine the number of points corresponding to this variable. Repeat this for each of the remaining variables. Add up the total number of points determined by each variable. Find the total number of points on the total point axis and draw a straight line down to the 5-year probability of recurrence axis. This provides the risk of recurrence within 5 years of initial treatment for MCC.

Figure 3:

A web-based MCC recurrence risk calculator is publicly available for patients and providers at merkelcell.org/recur. This calculator uses the multivariable risk regression model to generate recurrence risk estimates for individual patients. Chance of recurrence is an estimate within 5.7 years (the time of last event observed in this cohort). Curves are generated using a competing risk model as specified in the Methods section and graphically inverted to facilitate clinical interpretation.

Comparing a stage-only model to the multivariable model

The multivariable model was used to generate recurrence risk estimates for hypothetical patients within the same stage, but with different demographics and tumor characteristics (Figure 4). Recurrence risk within stage IIIA varied by up to a factor of 4 depending on non-stage factors: a stage IIIA patient with low-risk features may have a 5-year recurrence risk as low as 22% (78% recurrence-free) while that for a stage IIIA patient with high-risk features may rise to 82% (18% recurrence-free) (Figure 4). The 5-year recurrence risk estimate increased or decreased by more than 10% in 191 of 618 patients (31%) and in 77 of 179 p-IIIA patients (43%) when calculated based on the multivariable model versus using AJCC stage only (absolute difference). When AJCC stage alone was used to predict recurrence, the optimism-adjusted concordance index estimates were 0.66 (95% CI 0.63-0.69) at 2 years and 0.65 (95% CI: 0.62-0.68) at 5-years. The addition of other patient and tumor characteristics along with AJCC stage in the multivariable model significantly improved the concordance index by 0.04 (95% CI: 0.02-0.07, p<0.001) at 2 years (0.66 to 0.70) and by 0.03 (95% CI: 0.01-0.06, p<0.001) at 5 years over the stage-alone model (0.65 to 0.68).

Figure 4:

To demonstrate the significant heterogeneity of recurrence risk within one stage (for example, stage IIIA) the multivariable model was used to estimate MCC recurrence risk for low- and high-risk stage IIIA patients based on low- and high-risk non-stage factors (represented by the blue and orange curves, respectively). Recurrence risk for all cases of IIIA patients (n = 179) was estimated using the empirical cumulative incidence function as specified in the Methods section. Curves were then inverted to facilitate clinical interpretation.

Discussion

Summary

A relatively high proportion of patients with MCC (approximately 40%) experience recurrence of their cancer.⁴ Recently, stage-specific recurrence data have been published, and stage is an important prognostic variable.⁴ However, there is abundant evidence that recurrence risk is also affected by non-stage factors including sex, age, immune suppression status, and unknown primary status.^1,3,13–16 Indeed, patients with the same cancer stage may have as much as a 4-fold difference in recurrence risk depending on these non-stage factors. Furthermore, although 90% of recurrences occur within the first 3 years after initial treatment, it has been difficult to assess a given patient’s residual recurrence risk after diagnosis. Without a convenient risk prediction tool that integrates stage as well as non-stage variables and time since diagnosis, it is challenging to plan surveillance appropriately. Furthermore, NCCN guidelines regarding surveillance in MCC lack specificity, recommending imaging studies and follow-up “as clinically indicated”.¹⁷ Here, we present a web-based, easily-accessible tool for using this integrated model to calculate an individual patient’s MCC recurrence risk.

Findings in context of literature

Current MCC literature suggests that stage and other factors impact MCC prognosis. To date, many studies have demonstrated that higher stage / greater extent of disease is associated with poorer survival.^{15,16,18–21} Patient and tumor characteristics such as male sex, advanced age, known primary status, and immune suppression are significantly predictive of more aggressive disease.^1,3,13–16 For example, known primary status was predictive of significantly higher recurrence risk within stage IIIA patients (37% vs. 21% one-year recurrence risk for those with known vs. unknown primary, p = 0.03).⁴ However, the collective impact of non-stage variables on recurrence risk has not been previously quantified. Of note, in this multivariable model, age (HR: 1.10 per 10-year increase, p = 0.17) and unknown primary status (HR: 0.65, p = 0.064) did not have statistically significant hazard ratios for recurrence. However, this does not imply those factors are unimportant for recurrence risk. They were included in the model due to clinical relevance and the magnitudes of their estimated hazard ratios, which were appreciably greater or less than 1 and could thus materially impact risk predictions. Given the rarity of MCC, the sample sizes used to create this model were relatively small, and that may have limited our power to detect a statistically significant impact. Variables that are known to be clinically relevant are often included in risk prediction models, even if they do not quite achieve statistical significance.²²

Multiple studies have demonstrated that the highest risk period for disease recurrence is the first 3 years after diagnosis and initial treatment.^4,21,23 We sought to develop an online recurrence risk calculator which provides initial and residual recurrence risk estimation at any given point during follow-up (Figure 3). During subsequent follow-up, estimates of residual risk of recurrence help determine appropriate surveillance frequency with imaging and/or blood testing for recurrence. After 3 years of follow up, it will be clear that the vast majority of a patient’s recurrence risk has passed, and de-escalation of surveillance may be considered.

Nomograms and risk calculators are used widely as tools for estimating risk of cancer recurrence, mortality, and guiding treatment.²⁴ Our multivariable model improves risk estimation (concordance index 0.70) to an extent comparable to analogous tools in oncology.^25–27 Furthermore, the current calculator allows patients and providers to assess the decrease in risk during follow-up. This is particularly relevant in MCC because of rapid decreases in recurrence risk and the appropriateness of decreasing surveillance intensity.

Limitations

Although this multivariable risk calculator demonstrated good discrimination and calibration, there were relevant limitations. Some aspects of this cohort may not be representative of the current US population of patients with MCC. Our cohort included patients with advanced MCC (unresectable, stage III-IV disease), a small number of whom (28/618, <5%) received immunotherapy. Recurrence risks are likely higher for patients with advanced MCC treated prior to 2017 who received chemotherapy, as immune therapy was not yet widely available. We anticipate that as adjuvant immunotherapy trials are reported (including ADMEC-O, ADAM, and STAMP) ^{28 29,30} practice patterns may change. Therefore, for patients presenting with nodal or distant metastatic disease, the estimates from this study may not reflect the risk of recurrence for similarly-staged patients receiving treatment with immune therapy. Also, a greater proportion of patients in the Seattle cohort (78%) received some radiation therapy compared with the national population of patients with MCC (approximately 56%).^31,32 Because radiation therapy improves local control of MCC, it is possible that recurrence rates would be higher in a population with less use of or access to adjuvant radiation.^33,34 Given the need for some patients to travel long distances to the referral center, this cohort may include patients with higher socioeconomic status compared to the national average of MCC patients. Lastly, although Merkel oncoprotein antibody seronegative patients face increased recurrence risk^35,36, we did not include patients’ serologic status in our analysis. This is because, for many patients, baseline serologic status was not determined.

Future directions

In the future, inclusion of Merkel cell polyomavirus tumor status, MCPyV antibody titer levels, and/or circulating tumor DNA level could add predictive value to an MCC recurrence risk calculator.^35–37 Additionally, if an external data set containing MCC recurrence details becomes available -- such as via a multi-institution collaboration or another institution’s retrospective cohort -- the multivariable model’s discrimination and calibration should be tested. This would assess the model’s validity across MCC patient populations with different demographics and management regimens. Furthermore, as individual patients discuss the risks and benefits of adjuvant immunotherapy, this risk calculator should provide useful baseline information.

Conclusion

The MCC recurrence risk calculator based on this newly performed multivariable analysis advances recurrence risk estimation in two ways: (1) by providing more accurate recurrence risk estimates than using stage alone, and (2) by updating the recurrence risk estimates during the disease course based on time since initial treatment. This tool assists clinicians in educating patients regarding the likelihood their cancer will recur, and appropriately focusing resources on patients and time ranges in which recurrence risk is highest. Optimized surveillance should limit unnecessary costs while maximizing the chance of early recurrence detection.

Table 1:

Patient characteristics (N=618).

Variable	Value
Female sex	227 (36.7)
Age, years	69 (11 - 98)
<55 years	67 (10.8)
55-64 years	144 (23.3)
65-74 years	223 (36.1)
75+ years	184 (29.8)
8th Edition AJCC Stage
Pathologic Stage Total	498 (80.6)
P-I	183 (29.6)
P-II	47 (7.6)
P-IIIA	179 (29.0)
P-IIIB	63 (10.2)
P-IV	26 (4.2)
Clinical Stage Total	120 (19.4)
C-I	52 (8.4)
C-II	28 (4.5)
C-III	26 (4.2)
C-IV	14 (2.3)
Site of primary tumor
Extremity	243 (39.3)
Trunk	73 (11.8)
Head/neck	204 (33.0)
Unknown primary*	98 (15.9)
Size of primary tumor†
≤ 1 cm	202 (39.1)
1-2 cm	159 (30.8)
>2 cm	155 (30.0)
Immunosuppressed	82 (13.3)

Values are no. (%) or median (range);

^*Patients with an unknown primary typically present with disease in a lymph node and do not have any detectable skin lesion at diagnosis;

^†Based on 516 patients after excluding those with unknown primary (n = 98) and without a primary tumor size available (n = 4).