In this multicenter retrospective study, we firstly collected 4,089 OSCC patients from the Head and Neck Surgery Department of the Sun Yat-sen University Cancer Center (SYSUCC, Guangdong, China). The inclusion criteria of patients were as follows: (1) received pretreatment clinical assessment including tumor size and nodal involvement of physical and imaging examinations, (2) received curative-intent surgical resection, (3) received postoperative pathological confirmation, (4) follow-up time greater than 6 months. The patients were excluded according to the following criteria: (1) patients experienced distant metastasis at the time of diagnosis, (2) patients with previous history of OSCC, (3) subgroup with small sample size. After applying the criteria, 404 patients from SYSUCC between 2000 and 2016 were enrolled in this study and randomly split into the training cohort (n = 282) and internal validation cohort (n = 122) with 7/3 split ratio.

OSCC patients satisfying the aforementioned inclusion and exclusion criteria were also obtained from the Hospital of Stomatology of the Sun Yat-sen University between 2013 and 2018 (Guangdong, China). In total, 95 patients were designated as the external validation cohort. The screening process are shown in Figure 1 . All research procedures were approved by the Ethics Committee of the Sun Yat-sen University Cancer Center and Hospital of Stomatology of the Sun Yat-sen University. Informed consents for data collection and analysis were obtained from patients.

A total of 12 key variables were categorized into three data types for each OSCC case. Social demographic data included gender, age, radiotherapy history for head and neck cancer, and smoking history. Clinical data included primary site, physical examination-tumor size (PE-T), imaging examination-tumor size (IE-T), physical examination-nodal involvement (PE-N), and imaging examination-nodal involvement (IE-N). Imaging examination only included CT and MRI in our study. Pathological data included pathological tumor stage (P-T), pathological nodal involvement stage (P-N), and histologic grade. All the data were summarized in Table 1 . The clinical and pathological TN stage were classified according to the 7^th edition of the AJCC staging system for oral cancer. The endpoint of this study was overall survival (OS), which referred to the time interval from surgery to death or the last follow-up (12). The survival time of patients who were still alive at the last date of follow-up was given as censored data.

Kaplan–Meier analysis was used to estimate OS and detect intersections between the variables. The independent prognostic factors affecting OS were identified by univariate and multivariate Cox regression analyses. The qualified prognostic factors with significant differences in the univariate analysis were incorporated into the multivariate analysis. Stepwise regression was adopted to remove the non-significant factors, which ensured each variable in the resulting independent variable subset was significant to the dependent variable and the remaining variables were not multicollinear. The results were shown as hazard ratios (HR) with 95% confidence interval (CI). Then, an integrated nomogram was established to predict 1-, 3-, 5-year OS based on multivariate Cox proportional hazards regression model. Finally, the clinical–pathological model was validated internally and externally according to the TRIPOD statement (13).

The performances of the prognostic model were evaluated by various methods, involving the time-dependent receiver operating characteristic (ROC) curve and the value of the area under the ROC curve (AUC), Harrell’s concordance index (C-index), calibration curve, and decision curve analysis (DCA). ROC was used to assess the sensitivity and specificity of the model. C-index was determined to evaluate the model’s discriminative power between the predicted model and actual chance of experiencing the events (14). 1,000 bootstrap resamples were used to obtain the intervals of the C-index. The purpose of the calibration curve is to evaluate the agreement between the predictive values and observation values in the probabilities of 3- and 5-year survival of individuals. DCA was used to determine the net benefit of using the prognostic model at various threshold probabilities, which would be helpful to evaluate the actual needs for clinical decision-making (15). The total risk points of each patient were calculated according to the established nomogram. An optimal cut-off point was determined by the R package “maxstat” in the training cohort to classify patients as high-risk and low-risk groups.

All analyses were conducted using Python version 3.7.1 and R version 4.0.0 (R Foundation for Statistical Computing, Vienna, Austria). All statistical tests were two-sided with statistical significance defined as a p < 0.05. The Kolmogorov–Smirnov test was used to compare the distribution in the training and validation cohorts between different subgroups. Kaplan–Meier analysis was performed, and log-rank tests were used to determine the significance of the survival differences. In addition, Hosmer–Lemeshow tests were applied to evaluate the goodness-of-fit of the calibration curve, p > 0.05 represented good calibration (16).

The demographic and clinicopathologic characteristics of the training and validation cohorts are summarized in Table 1 . Of the 282 individuals in the training cohort, 68.09% of the patients were male, and patients over 60 years accounted for 35.82%. Only minority of patients (3.9%) had a radiotherapy history for head and neck cancer, and about half of patients (46.81%) had a smoking history. The 95 individuals in the external validation cohort were slightly younger with a lower prevalence of smoking. The median period of follow-up of the training cohort was shorter than that of the external validation cohort (25.6 vs. 42 months). The tongue was the most common primary tumor site in both the training and validation cohorts (50.35, 58.2, 66.31%, respectively). There was no statistically significant difference in the distribution of the features between the training and internal validation cohorts (p > 0.05), while the distribution of the primary tumor site, PE-T, and IE-N between the training and external validation cohorts was significantly different (p < 0.05) ( Table 1 ).

Kaplan–Meier analysis showed that smoking history, primary site, PE-T, IE-T, PE-N, IE-N, P-T, P-N, and histologic grade were significantly associated with OS (all p < 0.05), while gender, age, and radiotherapy history displayed non-significance (p > 0.05) ( Figure 2 ). All the available characteristics were included in the univariate analysis. There were statistically significant survival differences in the characteristics of primary tumor site, PE-T, IE-T, PE-N, IE-N, P-T, P-N, histologic grade in the univariate analysis ( Table 2 ). These significant variables in the Kaplan–Meier curves and the univariate analysis were included in the multivariate analysis to further screen out significant factors. In the multivariable stepwise regression analysis, PE-T (HR = 3.811; 95% CI, 1.210–12.004; p = 0.022), IE-T (HR = 4.135; 95% CI, 1.343–12.736; p = 0.013), P-N (HR = 1.834; 95% CI, 1.241–2.710; p = 0.002), and histologic grade (HR, 1.649; 95% CI, 1.083–2.511; p = 0.02) were significantly associated with OS in the training cohort ( Table 2 ). Meanwhile, P-T, P-N, and histologic grade were significantly related to the outcome in the pathological model (p < 0.05, Supplementary Table S1 ).

Based on the results of the multivariable Cox regression model, four independent variables were incorporated to develop a more accurate nomogram for optimizing personalized prognostic assessment and predicting 1-, 3- and 5-year OS ( Figure 3 ). Higher score was associated with a poor prognosis. The nomogram of the pathological model was shown in Supplementary Figure S1 .

The predictive accuracy between the clinical-pathological model and the pathological model has been compared. In the training cohort, the 1-, 3-, and 5-year AUC values of the clinical–pathological model for OS prediction were 0.687, 0.719, 0.722 ( Figure 4A ), which were superior to those of the pathological one (0.649, 0.707, 0.717, respectively) and AJCC TNM staging system (0.628, 0.668, 0.677, respectively), demonstrating excellent sensitivity and specificity for the clinical–pathological model ( Supplementary Figures S2A , S3A ). Similarly, the 1-, 3-, and 5-year AUCs in the internal validation cohort (0.775, 0.662, 0.687, respectively; Figure 4B ) and external validation cohort (0.918, 0.741, 0.787, respectively; Figure 4C ) of the clinical–pathological model also showed better discriminative ability compared with the pathological one and the AJCC TNM staging system ( Supplementary Figures S2B, C , S3B, C ).

The C-indices of the clinical–pathological model displayed better predictive performance than that of the pathological model in the training (0.664; 95% CI, 0.615–0.711 vs. 0.638; 95% CI, 0.592–0.683), internal validation (0.679; 95% CI, 0.609–0.75 vs. 0.655; 95% CI, 0.578–0.726) and external validation cohorts (0.755; 95% CI, 0.644–0.853 vs. 0.702; 95% CI, 0.621–0.778) ( Table 3 ). The clinical–pathological model exhibited superior discriminative power for OS prediction compared with the pathological model and the AJCC TNM staging system.

The calibration curves of 1-, 3- and 5-year OS and non-significant Hosmer–Lemeshow test demonstrated a good agreement between the prediction and observation values (p > 0.05) in the training ( Figures 4D, G, J ) and validation cohorts ( Figures 4E, F, H, I, K, L ), indicating that there was no deviation from the perfect fit. The calibration curves of pathological model were shown in Supplementary Figures S2D–L .

DCA analysis was conducted to evaluate the clinical value of our model. In the validation cohort, for predicted threshold probability between 30 and 70%, both the clinical–pathological and pathological models showed a positive net benefit for 3- and 5-year OS. Furthermore, the clinical–pathological model had a better net benefit for decision-making with the threshold probability within 30 and 50% illustrated by DCA ( Figures 4N, O ). Collectively, the threshold probabilities of the clinical–pathological model had better net benefits for predicting the 1-, 3- and 5-year OS in OSCC patients compared with the pathological one and the AJCC TNM staging system ( Figures 4M–O ).

Based on the individualized risk points of the nomogram, patients were divided into low-risk and high-risk groups in the training, internal, and external validation cohorts ( Figures 5A–C ). The optimal cut-off value was 79.08 for the clinical–pathological nomogram, and the Kaplan–Meier curves revealed that the high-risk group (total points > 79.08) was significantly correlated with a poor prognosis. The optimal point effectively distinguished populations of low-risk and high-risk, demonstrating a good prognostic classification for OSCC patients.