This study was conducted using nationwide data provided by the Health and Welfare Data Center (HWDC), established by the Ministry of Health and Welfare in Taiwan. The HWDC consolidates data gathered by the government from various sources, which is then de-identified and made available for research based on case-by-case approval17. Among the databases available in the HWDC, this study utilised three data sources: the National Health Insurance Research Database (NHIRD), which includes billing information on all National Health Insurance (NHI)-reimbursed examinations, medications, and treatments; the Taiwan Cancer Registry (TCR), which includes detailed staging and treatment information of cancer patients in Taiwan; and the Cause of Death database, which lists all death certificates issued in Taiwan. Reporting of NPC to the TCR started in 2009 with the long-form database, which included data on total radiation dose, modality, start and end dates of radiotherapy, timing of systemic and locoregional therapy (i.e. sequential or concurrent chemoradiotherapy) amongst other detailed information, requiring 114 fields in total for a single patient18,19. Notably, data on recurrence (including the date and site of recurrence) were also required elements, but updating beyond the initial registry entry was not mandatory. Nonetheless, quality assessments suggest that TCR ranks amongst the highest quality cancer registries not only in Asia but also worldwide19. All databases in the HWDC can be linked through a common but anonymised identifier. The latest edition of TCR available for analysis was 2015, while the latest edition of Cause of Death database was 2018.
This study received a certificate of exempt review from the Institutional Review Board of National Cheng Kung University Hospital. Requirement for informed consent was also waived. This research was performed in accordance with the Declaration of Helsinki.
We selected patients aged 20 years and above with a diagnosis of NPC (ICD-O-3 site: C11) and with pathologically confirmed invasive carcinoma (ICD-O-3M-codes: 8010, 8020, 8070, 8071, 8072, and 8082). Our inclusion criteria required upfront CCRT of at least 60 Gray via intensity-modulated radiotherapy or volumetric-modulated arc therapy. Patients with prior malignancy, two-dimensional or three-dimensional conformal radiotherapy, or radiotherapy alone were excluded. Patients receiving neoadjuvant chemotherapy were also excluded.
Patient covariates and outcome definition
We extracted data on age, sex, stage, Union for International Cancer Control/American Joint Committee on Cancer (UICC/AJCC) staging edition, treatment, and disease status at the last follow-up date from the TCR. Age was analysed as a continuous variable. Based on histopathological findings, we categorised the tumours as non-keratinising squamous cell carcinoma (ICD-O-3M-code: 8072), keratinising or unspecified squamous cell carcinoma (8070 or 8071), or other histologies including lymphoepithelial carcinoma, undifferentiated carcinoma, and unspecified carcinoma (8082, 8020, or 8010).
OS was calculated from the first day of radiotherapy to the day of death. The date of death was obtained from the Cause of Death database. Patients whose death records could not be found were considered alive and were censored on the last day of database records (31 December 2018). Disease-free survival (DFS) was defined as the time interval from the first day of radiotherapy to any recurrence; locoregional relapse-free survival (LRFS) and distant metastasis-free survival (DMFS) were defined as the time intervals from the first day of radiotherapy to locoregional or distant metastasis, respectively. DFS, LRFS, and DMFS were solely based on TCR data.
Designation of adjuvant PF chemotherapy and observation groups
To confirm adjuvant chemotherapy status, we required double confirmation from both the TCR and NHIRD. The TCR indicated whether adjuvant chemotherapy was administered. We searched the linked NHI reimbursement database within the window period of 7–90 days from the last day of radiotherapy for prescription of the following cytotoxic drugs: cisplatin, carboplatin, 5-fluorouracil, tegafur-uracil, epirubicin, mitomycin-c, doxorubicin, and methotrexate. Patients with both registry-documented adjuvant chemotherapy and prescription of both 5-FU and one of cisplatin or carboplatin (i.e. the PF regimen) within this period were included in the adjuvant PF chemotherapy group; conversely, patients whose registry records indicated a lack of adjuvant chemotherapy, along with an absence of any cytotoxic drug prescription (as stated above), were included in the observation group. The reason we chose to require double confirmation is to reduce treatment heterogeneity as much as possible. Patients coded as receiving chemotherapy in the TCR but not prescribed PF possibly received an alternate chemotherapy regimen, such as cisplatin-gemcitabine (not reimbursed in Taiwan) or tegafur-uracil. Patients coded as not receiving chemotherapy but prescribed with cytotoxic drugs may have been coded with inaccurate information due to severe delay in adjuvant chemotherapy, treatment at a different institution, or because chemotherapy was prescribed for a second primary malignancy diagnosed during the NPC treatment course. In either case, a discordant registry and reimbursement data indicates treatment deviation from the typical adjuvant PF course, and we exclude these patients to reduce treatment heterogeneity.
Baseline demographics and stage classification were compared using the chi-square test. Comparison of continuous variables was performed with the Kruskal–Wallis test or Student t-test.
We conducted univariate analysis by plotting Kaplan–Meier survival curves for previously defined endpoints and compared curves using the log-rank test. Multivariable Cox proportional hazards regression analysis was performed to estimate the independent effect of adjuvant chemotherapy.
Propensity score analysis was also conducted to assess the potential selection bias owing to imbalance of baseline factors resulting in decision to give adjuvant chemotherapy. The propensity score was created by fitting a multivariable logistic regression model including age, sex, histological subtype, clinical T classification, and N sub-classification; one-to-one matching was performed using nearest neighbour matching without replacement.
We performed landmark analyses to assess the effect of survival bias (immortal time bias). The typical adjuvant chemotherapy course is usually concluded within 6 months after the end of CCRT. Three separate analyses limited to patients surviving over 12, 18, and 24 months were performed.
All statistical analyses were conducted using SAS version 9.3 (SAS Institute, Cary, NC, USA), and R version 3.6.0 (R Foundation for Statistical Computing, Vienna, Austria). We calculated two-sided p-values with statistical significance defined at alpha = 0.05, along with 95% confidence intervals (CIs) to assess the precision of the estimates.