Liquid biopsy posttreatment surveillance in endemic nasopharyngeal carcinoma: a cost-effective strategy to integrate circulating cell-free Epstein-Barr virus DNA

We developed a Markov decision-analytic model to evaluate the cost-effectiveness of five surveillance strategies in NPC (Additional file 1: Fig. S1A): routine follow-up strategy, i.e., (1) routine clinical physical examination; routine imaging strategies, including (2) routine MRI plus CT plus bone scintigraphy (MRI + CT + BS); and (3) routine ¹⁸F-fluorodeoxyglucose PET/CT; cfEBV DNA-guided imaging strategies, including (4) cfEBV DNA-guided MRI + CT + BS; and (5) cfEBV DNA-guided PET/CT. The routine follow-up included history and physical examinations, complete blood counts, comprehensive metabolic panels, and nasopharyngoscopies. The routine imaging strategies included routine follow-up and imaging studies. The cfEBV DNA-guided imaging strategies included routine follow-up and cfEBV DNA tests, whose positive results would trigger further imaging studies. MRI imaged the head and neck; CT imaged the chest and abdomen; BS imaged the whole body; and PET/CT imaged from the skull base to midthigh. According to the previous study [20], the surveillance was scheduled as follows: (1) stage I: 2 follow-up visits in year 1; 3 visits in year 2; 2 visits in year 3 and 4; and 1 visit in year 5, (2) stage II: 2 visits in year 1; 4 visits in year 2; 2 visits in year 3 and 4; and 1 visit in year 5, (3) stage III: 4 visits in year 1 and 2; 3 visits in year 3; and 1 visit in year 4 and 5, and (4) stage IV: 4 visits in year 1; 5 visits in year 2; 3 visits in year 3; and 1 visit in year 4 and 5. We chose a 5-year follow-up duration because more than 90% of recurrence happened within this interval [21, 22].

We simulated the natural history of four hypothetical cohorts of 45-year-old asymptomatic nonmetastatic NPC patients with stage I, II, III, and IV disease after complete remission (CR) to the primary treatment. In each cohort, patients would move through the following health states: no evidence of disease (NED) after primary treatment, early- and advanced-stage recurrence (local relapse [LR], regional relapse [RR], or distant metastasis [DM]), salvage treatment for recurrence, NED after salvage treatment, and death (Additional file 1: Fig. S1B). In each 1-month cycle, the Markov model would accrue the costs and utilities as patients experienced different health states or events in a lifetime horizon. All future costs and utilities were discounted at a standard rate of 3% annually [23].

The primary outcomes included the cost, effectiveness, incremental cost-effectiveness ratio (ICER), and net health benefit (NHB) of each strategy. The ICER was defined as the incremental cost for an additional quality-adjusted life-year (QALY) gained by comparing two strategies. The willingness-to-pay threshold was set at $100,000/QALY [24, 25]. Strategies with ICERs less than this threshold were considered cost-effective. The NHB, which is explicit and robust when comparing multiple strategies, was calculated with the formula: effectiveness–cost/willingness-to-pay [26]. The strategy with the highest NHB was considered the most cost-effective at the given willingness-to-pay threshold. Additionally, we compared the average number of imaging studies required to detect one disease recurrence among the different surveillance strategies.

All analyses were conducted and reported in line with the recommendations of the Second Panel on Cost-Effectiveness in Health and Medicine and the Consolidated Health Economic Evaluation Reporting Standards [23, 27]. The institutional ethics committee approved the study and waived the requirement for informed consent given its retrospective nature (B2020-410). Tests were two-sided, and P < 0.05 was defined as statistically significant. The Markov model was constructed in TreeAge Pro (TreeAge Software, Williamstown, MA), and other statistical analyses were performed using R (version 3.6.1, http://​www.​r-project.​org).

The base-case estimates are presented in Additional file 1: Table S1 [3, 4, 16, 20, 23, 28‐56]. Monthly LR, RR, and DM probabilities were fitted by parametric survival models using the patient-level real-world data of an NPC cohort containing 10,097 nonmetastatic NPC patients (median follow-up time, 67.3 months) collected from the NPC-specific database in our institution; detailed information about the cohort and the database are presented in Additional file 1: Table S2 and Additional file 2: Supplementary Methods [38, 57‐62]. The fitness of exponential, Gompertz, log-logistic, log-normal, and Weibull distributions were evaluated [63]. Specifically, the log-normal distribution was applied to the LR data, while the Gompertz distribution was applied to the RR and DM data, as they demonstrated the best fit based on the Akaike information criterion [63, 64]. The age-specific background mortality rate was generated from the China life table (Additional file 1: Table S3) [46]. The starting age of 45 in the base-case analysis was the median age of the 10,097 patients. We assumed that undetected early-stage LR, RR, and DM had monthly probabilities of 10%, 10%, and 20%, respectively, to progress to advanced-stage recurrence based on the values utilized in a published study [20]. Other probabilities and clinical utilities for different health states were derived from published studies and, in few cases, from expert opinion if published data were unavailable (Additional file 1: Table S1). The literature search and data extraction are detailed in Additional file 2: Supplementary Methods. All transition probabilities were transformed into monthly scales via the declining exponential approximation of life-expectancy equation: μ = − 1/t × ln(S) [65, 66].

To evaluate the ability of the Markov model to accurately describe the disease processes of NPC, we validated the model in two aspects. First, we compared the model-simulated disease recurrence patterns (LR, RR, and DM) against the real-world recurrence patterns in the NPC cohort of 10,097 patients. The model-simulated recurrence was generated from the Markov model using the base-case parameters, while the real-world recurrence was calculated using the Kaplan-Meier method. Second, we compared the model-predicted overall survival under the five surveillance strategies in this study with the real-world observed overall survival in the NPC cohort (Kaplan-Meier method). Although the patients in real-world clinical practice might be prescribed various follow-up modalities and intervals different from this study, the similarity between the model-predicted and observed survival could partially confirm the model’s ability to reproduce the natural history of NPC patients because the recurrence patterns in the model were supposed to be similar to those in the real-world cohort. To better mimic the real-world circumstance consisting of various follow-up methods, we calculated the model-predicted overall survival using the average of the overall survival of the five surveillance strategies in the model.

Costs were derived from the 2019 Medical Insurance Administration Bureau of Guangzhou, China, or published studies from the societal perspective (Additional file 1: Table S1). Direct medical costs included costs of follow-up and salvage treatment. The costs assigned to a false-positive cfEBV DNA test or routine follow-up included the costs of additional imaging studies, while those assigned to a false-positive imaging study included the costs of nasopharyngoscopy- or CT-guided biopsy, pre-procedure labs, and pathology processing. Patients with early-stage LR would undergo endoscopic nasopharyngectomy (ENPG, 70%) or reirradiation (30%) [30], while patients with advanced-stage LR would undergo reirradiation. Patients with RR would undergo neck dissection, while patients with DM would receive salvage chemotherapy. The societal costs included costs of transportation, accommodations, meals, and wage loss related to patients’ and caregivers’ time off work [23, 53‐55]. All costs were inflated to 2019 using the medical care component of the Consumer Price Index in China and were converted into 2019 US dollars (exchange rate, 1.00 USD = 6.91 CNY).

One-way sensitivity analyses were applied to parameters over plausible ranges derived from published literature or a deviation of 20% from the base-case values (Additional file 1: Table S1) [53, 64]. Probabilistic sensitivity analyses (PSA) were conducted using 10,000 Monte Carlo simulations, where all parameters were varied simultaneously based on specific probability distributions (beta distribution for probabilities and utilities; gamma distribution for costs). Cost-effectiveness acceptability curves were constructed based on the strategies’ NHB generated from PSA. Additionally, two scenario analyses were performed by varying the surveillance arrangements: every 3 months in years 1–2 and every 6 months in years 3–5 according to the Radiation Therapy Oncology Group (RTOG) [67, 68]; every 2 months in year 1, every 4 months in year 2, and every 6 months in years 3–5 according to the NCCN guidelines for head and neck cancer (Version 1.2021) [19].

The Markov model-simulated LR-free survival, RR-free survival, and DM-free survival showed good agreement with the real-world disease recurrence patterns (Additional file 1: Fig. S2). In addition, the absolute differences between the model-predicted 1-year, 3-year, and 5-year overall survival and those of the real-world observation were within 2% in all disease stages, which indicated that the patient’s outcomes were comparable between the Markov model and the real-world observation (Additional file 1: Table S4). Collectively, these validation results demonstrated that the Markov model could accurately describe the disease processes of NPC.

The results of the base-case analysis are summarized in Table 1 and Additional file 1: Table S5. With the assistance of cfEBV DNA, cfEBV DNA-guided imaging strategies were considerably more cost-effective than routine imaging strategies across all disease stages. Specifically, compared with cfEBV DNA-guided MRI + CT + BS, the ICERs of routine MRI + CT + BS or routine PET/CT were both exceeding the willingness-to-pay threshold of $100,000/QALY, ranging from $179,508/QALY in stage IV to $1,443,575/QALY in stage I (Additional file 1: Table S5). Similarly, the ICERs of routine imaging strategies compared with cfEBV DNA-guided PET/CT also went far beyond the willingness-to-pay threshold, ranging from $262,724/QALY in stage IV to $3,882,176/QALY in stage I, and the routine MRI + CT + BS was dominated (less effective but more costly than another strategy) by cfEBV DNA-guided PET/CT in stage IV (Additional file 1: Table S5). Moreover, the NHBs of cfEBV DNA-guided imaging strategies were also greater than those of routine imaging strategies (Table 1). For example, the NHBs of cfEBV DNA-guided MRI + CT + BS and PET/CT were 12.066/QALY and 12.061/QALY in stage II, respectively, larger than those of routine MRI + CT + BS (12.029/QALY) and PET/CT (11.979/QALY).

Next, we scrutinized the most cost-effective strategies in each stage of asymptomatic NPC patients (Table 1). For patients with stage I disease, cfEBV DNA-guided imaging strategies were only associated with a 0.01 gain in QALY compared with routine clinical physical examination. This translated to relatively large ICERs for cfEBV DNA-guided MRI + CT + BS ($119,368/QALY) and cfEBV DNA-guided PET/CT ($518,871/QALY), both surpassing the willingness-to-pay threshold. Therefore, routine follow-up was adequate for stage I NPC patients. However, in stage II and III, cfEBV DNA-guided MRI + CT + BS was the most cost-effective strategy, with ICERs of $57,308/QALY and $46,860/QALY, respectively, compared with routine clinical physical examination. Intriguingly, for patients with stage IV disease, cfEBV DNA-guided PET/CT became the most cost-effective, with an ICER of $62,269/QALY compared with cfEBV DNA-guided MRI + CT + BS. Likewise, the results of NHBs supported the same strategies mentioned above; the strategies with the highest NHBs were routine clinical physical examination in stage I (NHB, 13.166/QALY), cfEBV DNA-guided MRI + CT + BS in stage II (12.066/QALY) and stage III (11.392/QALY), and cfEBV DNA-guided PET/CT in stage IV (9.831/QALY). The preferred strategy in each stage is summarized in Fig. 1.

To further clarify the superiority of cfEBV DNA-guided imaging strategies over routine imaging strategies, we compared their average number of imaging studies required to detect one disease recurrence (Table 2). While routine imaging strategies resulted in a 5-year total of 97,493 to 118,328 imaging studies being performed per 10,000 patients, the cfEBV DNA-guided imaging strategies only required 19,676 to 25,893 imaging studies to detect a comparable number of recurrences as routine imaging strategies. Ultimately, compared with routine imaging strategies, cfEBV DNA-guided imaging strategies only required approximately one quarter of the imaging studies to detect one recurrence. For example, in stage II, routine MRI + CT + BS and routine PET/CT needed 121 and 106 imaging studies to detect one recurrence, respectively, while the guidance of cfEBV DNA decreased the number of imaging studies to 25 and 23, respectively. Therefore, cfEBV DNA-guided imaging strategies were the most favorable option for stage II to IV NPC patients. However, for stage I NPC patients, although the total imaging studies of cfEBV DNA-guided imaging strategies were remarkably fewer than routine imaging strategies, they still required a fair number of studies per recurrence detection (cfEBV DNA-guided MRI + CT + BS, 94; cfEBV DNA-guided PET/CT, 89) due to low recurrence probabilities, suggesting that, despite the guidance of cfEBV DNA, imaging was still unfavorable for them. This result was consistent with the cost-effectiveness analysis mentioned above that routine clinical physical examination was enough for patients with stage I disease.

To evaluate the results’ robustness, sensitivity analyses were performed on all probabilities, utilities, and costs. In one-way sensitivity analyses, cfEBV DNA-guided imaging strategies were consistently more cost-effective than routine imaging strategies across the range of all parameters (Additional file 1: Fig. S3–6). However, the optimal surveillance strategy in each stage was sensitive to several parameters when compared to the next more costly strategies, with key parameters displayed in Fig. 2A–D.

In stage I, the optimal strategy was sensitive to the LR probabilities and the specificity of routine clinical physical examination for RR (Fig. 2A). When increasing the LR probabilities to 2.1% or decreasing the specificity of routine clinical physical examination for RR to 65.9%, the preferred strategy would switch from routine clinical physical examination to cfEBV DNA-guided MRI + CT + BS. Nevertheless, all ICERs remained higher than $87,000/QALY when varying the parameters. In stage II–IV, the cost of PET/CT was the principal determinant of the ICERs (Fig. 2B–D). Specifically, if the PET/CT's cost was reduced to $1221, the cfEBV DNA-guided PET/CT would become the optimal strategy in stage III. Conversely, although the cost of PET/CT also had the greatest impact on the ICERs in stage II and IV, varying it did not change the optimal surveillance strategies in the base-case analysis.

The PSA results are illustrated in the cost-effectiveness acceptability curves shown in Fig. 3A–D. With the willingness-to-pay threshold varying from 0 to $200,000/QALY, the probabilities of routine imaging strategies being the optimal strategies were almost zero in stage I–III and less than 10.0% in stage IV. At a willingness-to-pay threshold of $100,000/QALY, routine clinical physical examination was the most cost-effective in 78.2% of the simulations for stage I patients; cfEBV DNA-guided MRI + CT + BS was the most cost-effective in 88.4% and 56.3% of the simulations for stage II and III patients, respectively; and cfEBV DNA-guided PET/CT was the most cost-effective in 96.2% of the simulations for stage IV patients.

To examine the cost-effectiveness of cfEBV DNA-guided strategies under different follow-up arrangements, we performed two scenario analyses using the recommended arrangements from the RTOG and NCCN guidelines (Additional file 1: Table S6) [19, 67, 68]. For both scenarios, cfEBV DNA-guided imaging strategies were consistently more cost-effective than routine imaging strategies. The optimal surveillance strategy in each disease stage was in line with the base-case results.