Study population
UMBTs were defined as a group of bone tumors in addition to osteosarcoma, chondrosarcoma, Ewing sarcoma and other specific bone tumors, which were provided by the ICCC-3 and encoded in the SEER database. By using the SEER*Stat software (Version 8.3.5, National Cancer Institute, Bethesda, MD, USA), the clinical information of UMBT patients from 1973 to 2016 was obtained from SEER database, which covers 30% US population [6, 7]. According to the ICD-O-3, the codes for selected histologic types were 8000–8005, 8800–8801, and 8803–8805, and the primary site codes were C40.0–40.3, C40.9–41.4, and C41.8–41.9. The anatomic site record was “bone and joints”. This study was a retrospective design with duration of over 40 years (1973–2016), and all the enrolled patients were from the USA population.
Inclusion and exclusion
In our study, the inclusion criteria were as follows: 1) Confirmed diagnosis of UMBT with positive histology; 2) the diagnosis was acquired at a living status; 3) active follow up was confirmed for each patient; 4) definite survival months. While the enrolled patients were further excluded for at least one of the following reasons: 1) incomplete information of tumor sites; 2) lack of surgical information; 3) patients under 18 years old; 4) the survival month was 0; 5) number of patients from Alaska < 11(please note we did not give the exact number to protect patient privacy). Notably, according to the provided information of the SEER database, tumor sites were divided into the following three subgroups for further analysis: limbs (including upper and lower limbs, and associated joints), and the spine (including vertebral column/pelvis, sacrum, coccyx, and associated joints), and others (including mandible/skull, face and associated joints/rib, sternum, clavicle and associated joints).
Baseline information
The collected clinicopathological factors included: age, race, sex, year of diagnosis, region (southwest/pacific coast/northern plains/east), primary tumor site, pathological grade, tumor laterality, tumor extent (localized/regional/distant), primary tumor surgery, radiotherapy and chemotherapy, tumor size, initial tumor diagnosed, hispanic/non-hispanic patients, marital status, and annual family income. The primary endpoints were cancer-specific survival (CSS) and overall survival (OS). CSS was defined as the interval from the first day to the death date due to UMBT or the end of follow-up, while OS was defined as the duration from the first day to the date of all-cause death or the end day of follow-up. Figure 1 showed the flow chart of patients collecting process: in short, 545 UMBT patients were initially screened out from the case information list via SEER*Stat software, then 145 patients were further excluded for the following reasons: Age < 18 years old(n = 54), unknown tumor site (n = 27), incomplete surgical information (n = 17), 0 survival month of follow up (n = 7), and the patient from Alaska (n = 1) was also removed because of the limited number. Finally, 400 UMBT cases were enrolled in our study for further analysis.
Statistical analysis
The mean with standard deviation (SD) and/or range were used to describe the quantitative data, while counts and percentages for qualitative data. The cutoff values of continuous factors were identified optimally by using the x-tile software (Yale University, New Haven, CT, USA) [8]. Theoretically, the optimal cutoff value was generated based on the highest chi-square values calculated upon the division. Notably, the results showed that the binary cutoff values of CSS and OS were described as follows: 66 years old for age, and 105/108 mm for tumor size, respectively (Fig. 2).
Univariate analysis was then performed by the SPSS 22.0 (SPSS Inc., Chicago, IL, USA) to figure out the potential significant factors by log-rank test, and parameters with p < 0.05 were further subjected to multivariate Cox proportional hazard model to identify the independent determinants with estimated hazard ratios. Notably, since the information of patients who did not receive radiotherapy and chemotherapy was incomplete in the SEER database (“None/Unknown” for radiotherapy, and “No/Unknown” for chemotherapy), these two factors were not subjected into survival analysis. Kaplan-Meier curve of independent factors was drawn by the GraphPad Prism 7.0 (GraphPad Software, San Diego, CA, USA).
After identifying independent contributors by multivariate analysis, Nomogram was further developed by using R software (version 3.5.1) with rms package (available at: http://CRAN.R-project.org/package=rms). The developed nomogram can be used to predict the 2 and 5-year CSS and OS, respectively. Harrell’s concordance index (C-index) and calibration curves were generated to assess the predictive value of the nomogram by comparing with the actual proportions.
Besides, the clinical features between surgery and non-surgery groups were assessed by using student t and Pearson χ2 test or Fisher exact test. Propensity score matching (PSM) method (R with MatchIt package, available at: http://CRAN.R-project.org/package=MatchIt) was utilized to balance the impact caused by the unmatched bias. The matching ratio was designated as 1:1 with caliper of 0.05 for surgery and non-surgery groups. After PSM, the effects of surgical treatments on CSS and OS were further evaluated among patients receiving either radiotherapy or chemotherapy, respectively. The related death rate curves were analyzed and depicted by using the GraphPad Prism 7.0 (GraphPad Software, San Diego, CA, USA), with p < 0.05 being regarded statistically significant.