2.1 Data source and study population

We conducted a retrospective cohort study using the NPC-specific database from the Big-data Intelligence Platform of Sun Yat-Sen University Cancer Center (Guangzhou, China). This database is a patient-level research system enabling real-time organization, integration, and updating of medical records automatically from several clinical systems based on well-designed data model and algorithms. It has been applied to various clinical studies [14-17]. Details of this platform were as previously described [15].

We retrieved the data of 10,058 patients with non-metastatic, biopsy-confirmed World Health Organization (WHO) type II/III NPC diagnosed cases between April 2009 and December 2015 from the NPC-specific database. All patients received routine pretreatment evaluations and were restaged according to the Union for International Cancer Control/American Joint Committee on Cancer (UICC/AJCC) Eighth edition staging system [18]. Details on the clinical workup and restaging are presented in the Supplementary Methods. The institutional ethics committee approved the study protocol and waived the requirement for informed consent given the study's retrospective nature. The study's authenticity has been validated by uploading the key raw data onto the research data deposit (RDD) public platform (http://www.researchdata.org.cn, approval RDD number: RDDA2018000934).

2.2 Treatment

All patients received radical IMRT as primary treatment. During the study period, our institutional treatment guidelines recommended IMRT alone for stage I patients and platinum-based CCRT ± induction/adjuvant chemotherapy for stage II-IVA patients, based on the recommendations of the updated National Comprehensive Cancer Network (NCCN) guideline for head and neck cancer version 1.2020 [19]. Reasons for deviation of patient treatment from our institutional treatment guidelines mentioned above included age, patient's refusal of treatment, or organ dysfunction suggestive of intolerance to treatment. Salvage treatments, including re-radiation, chemotherapy, and surgery, were provided during recurrence or persistent disease despite the above-mentioned treatments. The chemoradiotherapy protocols are detailed in the Supplementary Methods.

2.3 Follow-up and outcomes

The patients were regularly followed-up every 3 months during the first 2 years, every 6 months during the first 3–5 years, and annually thereafter. At each visit, physical examinations, plasma EBV DNA load, and fiberoptic nasopharyngoscopy were routinely performed. MRI of the nasopharynx and neck, chest radiography, abdominal sonography, bone scan, or PET/CT were repeated annually or when clinically suspected recurrence occurred. The follow-up duration was calculated from the date of diagnosis to either date of death or last follow-up. The date of last follow-up was October 2019.

The primary endpoint was overall survival (OS), defined as the time from diagnosis to death from any cause. The secondary endpoints included disease-free survival (DFS), the time from diagnosis to tumor recurrence at any site or death from any cause, whichever came first; locoregional relapse-free survival (LRRFS), the time to locoregional relapse; distant metastasis-free survival (DMFS), the time to metastasis; and NPC-specific survival (NPC-SS), the time to death from NPC. Patients without evidence of any events of interest were censored at the time of the last follow-up.

2.4 Statistical analysis

Continuous variables were transformed into categorical variables. Plasma EBV DNA load was categorized by every 10-fold increase as previously described [20-22]. The cut-off values for other laboratory variables were determined using maximally selected rank statistics based on OS, a widely used method that generates cut-off values with the most significant log-rank statistics [23-25]. The handling of cut-off values is detailed in the Supplementary Methods.

2.5 Conditional survival

CS denotes the probability of survival for an additional $y$ years given an existing survival time of $x$ years, which is calculated as $CS(y|x)=S(x+y)/S(x)$. The concept of CS can extend to multiple endpoints, deriving conditional OS (COS), conditional DFS (CDFS), conditional LRRFS (CLRRFS), conditional DMFS (CDMFS) and conditional NPC-SS (CNPC-SS) [1, 3, 26]. For example, the 3-year CDFS at 5 years demonstrates the probability of being disease-free for an additional 3 years given an existing disease-free time of 5 years.

We evaluated CS estimates for all endpoints mentioned above. Crude CS estimates were assessed using the Kaplan-Meier method. Adjusted 3-year conditional risk (1 minus CS estimates) stratified by prognostic factors of each endpoint was calculated after adjusting for other covariates using IPW, an algorithm used to balance covariates without sacrificing sample size and statistical power [27-29]; clinicopathological variables (i.e., tumor [T] stage, node [N] stage, EBV DNA, age, sex, histology, smoking, alcohol consumption, family history of NPC, lactate dehydrogenase [LDH], albumin [ALB], C-reactive protein [CRP] and hemoglobin [HGB]) that were statistically significant in IPW-adjusted log-rank tests (P < 0.050) were considered to be the prognostic factors of the endpoints. Considering that the median follow-up time of the primary cohort was 67.2 months, we chose the 3-year conditional risk because it permitted the estimation of conditional risk given existing survival times of 0-5 years, which has been widely used in studies of CS in other malignancies with similar median follow-up times [26, 30-33]. Absolute standardized difference (ASD) and chi-square tests were employed to determine the balance of baseline covariates after IPW adjustment. ASD values less than 0.1 indicated negligible imbalance [29, 34]. Then, patients were categorized in a competing risk framework according to their first events (i.e., local relapse [LR], regional relapse [RR], bone metastasis [Bone-M], liver metastasis [Liver-M], lung metastasis [Lung-M], NPC-specific death [NPC-SD] and other cause-specific death [OC-SD]). The conditional risk of each competing event was calculated using the cumulative incidence function.

2.6 Development and validation of prediction models

To reduce overfitting, our primary cohort was split into a derivation cohort and an internal validation cohort according to the diagnosis time. Models were constructed using cause-specific hazard models. Variables with P < 0.100 in univariate analysis were subjected to multivariate analysis to identify independent predictors for the models using backward selection with the Akaike information criterion. The proportional hazards assumption was verified based on the Schoenfeld residuals [35].

The models were validated internally and externally using the Harrell's C-index and calibration plots, and they were compared with the current 8^th edition UICC/AJCC staging system. The bias-corrected C-index was obtained using 1000 bootstraps resamples in the derivation cohort. The models were externally validated in an independent cohort of 601 non-metastatic, biopsy-confirmed WHO type II/III NPC patients diagnosed between February 2012 and July 2015 from Wuzhou Red Cross Hospital receiving radical IMRT with or without chemotherapy to evaluate the models’ general applicability. The date of last follow-up was October 2019, and other information related to this cohort is detailed in the Supplementary Methods.

2.7 Handling of missing data

For the primary cohort of 10,058 patients, including the derivation cohort and the internal validation cohort, missing data were considered missing completely at random. Therefore, the complete-case analysis was performed, resulting in a complete-case cohort of 9302 patients. The patient characteristics were well balanced between the original cohort and the complete-case cohort (Supplementary Table S1). It is worth noting that the analyses for the crude CS estimates and the conditional risk of competing events used the primary cohort of 10,058 patients because no missing values were observed during these analyses, while other analyses utilized the complete-case cohort.

For the external validation cohort, we utilized the multiple imputations by chained equations (MICE) with the predictive mean matching method to imputed missing values [36-39]. To guarantee the robustness of the validation, 1000 imputations with 30 iterations were carried out to yield the pooled Harrell's C-index and calibration plots based on Rubin's rules [36, 40, 41]. All tests were two-sided, and P < 0.050 was defined as statistically significant. Analyses were performed using the R software, version 3.6.1 (R Foundation for Statistical Computing, Vienna, Austria). Detailed statistical considerations are presented in the Supplementary Methods.

3.1 Patient characteristics

A total of 10,058 eligible patients were investigated. All patients received IMRT, and 8839 (87.9%) of the patients received additional chemotherapy (stage I: 0 [0%], stage II: 1414 [79.7%], and stage III–IV: 7425 [96.0%]). The median follow-up time was 67.2 months (interquartile range, 53.9-83.0 months). Among them, 1059 (10.5%) patients developed locoregional relapse, 1306 (13.0%) patients developed distant metastasis, and 1581 (15.7%) patients died (1365 [13.6%] were NPC-SD). Patient characteristics are presented in Supplementary Table S2.

3.2 Crude CS estimates of multiple endpoints

We first investigated how crude CS estimates of COS, CDFS, CLRRFS, CDMFS, and CNPC-SS changed with time. The CS probabilities improved with each additional year of existing survival time for the abovementioned endpoints (Figure 1A–D, Supplementary Figure S1A). Contrary to the ever-decreasing traditional survival estimates, the 1-, 3- and 5-year CS probabilities gradually increased as a function of the existing survival time (Figure 1E–H, Supplementary Figure S1B). Specifically, the 5-year COS, CDFS, CLRRFS, CDMFS and CNPC-SS increased from 86.2%, 78.1%, 89.8%, 87.3%, and 87.6% at diagnosis (i.e., an existing survival time of 0 year) to 87.3%, 87.7%, 94.4%, 96.0%, and 90.1%, respectively, given an existing survival time of 3 years since diagnosis (Supplementary Table S3). In addition, the 3-year CDFS, CLRRFS, and CDMFS increased considerably in the first 3 years but plateaued thereafter, indicating that the patients’ prognosis improved primarily in the first 3 years (Figure 1F–H). In contrast, the 3-year COS and CNPC-SS improved continuously without plateauing until beyond the fifth year (Figure 1E, Supplementary Figure S1B). Interestingly, the almost overlapping curves of “COS at 1 year” and “OS” in Figure 1A denoted that the COS with a 1-year existing survival time barely improved compared with the baseline survival estimates at diagnosis. Likewise, CNPC-SS at 1 year also increased minimally against the baseline estimates (Supplementary Figure S1A).

3.3 IPW-adjusted conditional risk stratified by prognostic variables

To further understand how prognostic factors of NPC influenced conditional risks, we calculated the 3-year conditional risk separately stratified by the prognostic factors of each endpoint after adjusting for other covariates using IPW. The IPW algorithm-generated well-balanced cohorts with comparable baseline covariates between subgroups (Supplementary Figure S2), and the ASDs of covariates considerably decreased after IPW adjustment, with values all below or near the thresholds indicating negligible imbalance (Supplementary Table S4). The prognostic factors independently associated with each endpoint in IPW-adjusted log-rank tests are presented in Supplementary Table S5. The number at risk of each subgroup is listed in Supplementary Table S6.

For T stage, N stage, and EBV DNA, the disparities in the 3-year conditional risk of all-cause of death and disease failure among each subgroup were notable at diagnosis, but they gradually tended to level with time (Figure 2A–F), and the same trend was observed for the stratified conditional risk of locoregional relapse, distant metastasis, and NPC-SD (Supplementary Figure S3A–H). This trend occurred due to the conditional risk for patients with an initially poorer prognosis (e.g., T4, N3, and EBV DNA ≥ 1000 × 10³ copies/mL), which decreased more substantially than for patients with an initially better prognosis. For instance, the 3-year conditional risk of disease failure for T4 decreased from 20.1% at diagnosis to 7.1% by the fifth year of disease-free survival, while those with T1 only decreased from 7.6% to 6.0% (Figure 2D).

Other prognostic factors for each endpoint (i.e. sex, histology, LDH and HGB for all-cause death, disease failure, distant metastasis and NPC-SD; smoking for disease failure and locoregional relapse; and ALB and CRP for all-cause death) showed similar patterns as mentioned above (Supplementary Figure S4A–T). However, age demonstrated enlarged disparities over time in the 3-year conditional risk of all-cause death, disease failure, and NPC-SD between subgroups (Figure 3A–C). Interestingly, this was because the conditional risk of recurrence or death for the older group (initially poorer prognosis) reduced less than that for the younger group (initially better prognosis), which was opposite to other factors. A possible reason was that older patients might experience a higher prolonged risk of recurrence or death due to increased comorbidity burdens, decreased performance status, and reduced tolerance to treatments [17, 42-45].

3.4 Establishment of prediction models for the conditional risk of site-specific recurrence

To clarify the conditional risk of specific events, the patients were further categorized according to their first events in a competing risk framework, and the cumulative risk of each competing event was estimated (Supplementary Figure S5). We then assessed how the conditional risk of each competing event changed with the existing disease-free time. The 3-year conditional risk of LR, RR, bone-M, lung-M, and liver-M decreased along with the existing disease-free time after diagnosis, and the decrease in metastases was more remarkable than that in LR or RR (Figure 4). In contrast, the 3-year conditional risk of NPC-SD remained nearly constant, while that of OC-SD gradually increased. The 1-year and 5-year conditional risk of specific events displayed similar trends to those mentioned above (Supplementary Figure S6A and B). Notably, the 3-year conditional risk for OC-SD surpassed the risk of bone-M, lung-M, and liver-M, given a 3-year disease-free time, while that for OC-SD exceeded the risk of RR and LR at 4-year and 5-year disease-free times, respectively (Figure 4).

Given that different sites showed unique conditional risk patterns, we established competing risk models for LR, RR, Bone-M, Lung-M, and Liver-M, respectively. Patient characteristics of the derivation cohort are listed in Supplementary Table S7. Variables with P < 0.100 in univariate analyses (Supplementary Table S8) were entered into multivariate analyses to select the independent predictors. The final prediction models are detailed in Table 1. To facilitate user-friendly access, the models were then integrated into a web-based application available at https://cr-npc.yiducloud.com.cn, which allows individualized conditional risk estimation of site-specific recurrence by entering patients’ clinicopathological information and existing disease-free time.

3.5 Performance of the prediction models and external validation

In the derivation and internal validation cohorts, the C-indices for the models to predict LR, RR, Bone-M, Lung-M and Liver-M were 0.656 and 0.656, 0.667 and 0.676, 0.742 and 0.709, 0.681 and 0.669, and 0.711 and 0.708, respectively, which were all significantly higher than those for the current staging system (P < 0.001), with respective values of 0.613 and 0.628, 0.581 and 0.578, 0.634 and 0.637, 0.647 and 0.616, and 0.625 and 0.615 (Table 1). The calibration plots indicated good agreement between the model-predicted and observed survival estimates in both cohorts (Supplementary Figure S7A, B, D, E, G, H, J, K, M, and N).

To investigate the general applicability of the models, they were further validated externally with an independent cohort of 601 NPC patients from Wuzhou Red Cross Hospital (median follow-up time = 61.3 months [interquartile range, 49.9-67.9 months]), with patient characteristics presented in Supplementary Table S7. Intriguingly, the C-indices for LR, RR, Bone-M, Lung-M and Liver-M were 0.672, 0.736, 0.754, 0.663, and 0.721, respectively, which significantly outperformed the current staging system (P < 0.050), with C-indices of 0.642, 0.673, 0.659, 0.618 and 0.612 (Table 1). The calibration plots also demonstrated good agreement between the predictions and observations (Supplementary Figure S7C, F, I, L, and O).