Prognostic and predictive significance of tumor length in patients with esophageal squamous cell carcinoma undergoing radical resection

This study was approved by the institutional review board of Zhejiang Cancer Hospital and the need for individual patient consent was waived. The study was conducted with data collected from a prospectively collected database for esophageal cancer. Between January 2003 and December 2010, 1613 consecutive cases were surgically treated at the Department of Thoracic Surgery of Zhejiang Cancer Hospital. Because an institutional electronic medical record system was used in our hospital since January 2003, this date was chosen as the starting date for the study. A total of 1435 patients with esophageal squamous cell carcinoma after resection with curative intent were included in this study (Fig. 1). Among 47 patients excluded because of incomplete resection, 35 patients had macroscopic residual disease (R2 resection) and 12 patients had microscopic disease (R1 resection: positive proximal resection margin in nine cases and positive distal resection margin in three cases). Seventeen patient with previous cancer history (gastric cancer in eight cases, lung cancer in four cases, laryngeal caner in three cases, breast cancer in one cases and malignant lymphoma in one cases) were excluded. Of 12 patients excluded because of synchronous cancer, seven patients had synchronous gastric cancers, three patients had synchronous hypopharyngeal cancers, one patient had a synchronous laryngeal cancer, and one patient had synchronous leukemia. Sixteen patients with non-squamous carcinoma (adenocarcinoma in six cases, adenosquamous carcinoma in four cases, small cell carcinoma in four cases, and carcinosarcoma in two cases) were also excluded. Because neoadjuvant therapy may influence postoperative pathological staging and tumor length, patients with neoadjuvant therapy were excluded. All of these 1435 patients received preoperative evaluations including endoscopy with biopsy, barium swallow examination, computerized tomography of the chest and upper abdomen, and ultrasound of the neck. Pulmonary and cardiac function tests were routinely performed to assess medical operability. Recurrent laryngeal nerve palsy and the presence of clinical supraclavicular or cervical nodal involvement were considered a contraindication for surgery. Histological diagnosis of each of the patients was established before treatment. Written informed consents were obtained from all patients before surgery.

Three surgical approaches were commonly used: Ivor Lewis procedure, cervico-thoraco-abdominal approach (Mckeown prodcedure), and left thoracotomy approach (Sweet procedure). Ivor Lewis procedure and Sweet procedure with anastomosis in the chest apex were usually performed when the tumor located in the lower and middle segment of the esophagus. When the tumor located in the middle or upper segment of the esophagus, Mckeown procedure with anastomosis in the left neck was mainly conducted (Fig. 1). Meanwhile, the choice of surgical procedure also depended on surgeons’ preferences. Two-field (mediastinal and upper abdominal) lymph node dissection was routinely performed for all patients. The extent of mediastinal lymph node dissection included all nodal tissue associated with esophagus in the chest from the superior mediastinal nodes and nodes along both recurrent laryngeal nerves to the hiatus. The extent of upper abdominal lymph node dissection included the paracardial, lesser curvature, left gastric, common hepatic, celiac, and splenic nodes. Three-field (cervical, mediastinal and upper abdominal) lymph node dissection was not routinely performed. However, this procedure was also performed selectively by surgeons depending on their preference. The extent of cervical lymph node dissection included supraclavicular and cervical paraesophageal nodes.

After surgical resection, the esophageal specimen was opened longitudinally from proximal to distal, extending this incision along greater curve of stomach if attached. The anatomical locations of the removed nodes were labeled by the operating surgeon. All specimens were fixed in 10 % formalin overnight, unpinned. and then sent to pathological examination. Tumor length was measured to the closest to 1 mm. In addition to tumor length, pathological details including histology type, differentiation, depth of invasion, lymph node status, vascular invasion, perineural involvement, the number of resected lymph nodes as well as proximal and distal surgical resection margin were reported. Circumferential resection margin was not routinely examined at our institution. Data from pathological reports were reviewed retrospectively. All patients were restaged based on the 7th edition of the American Joint Committee on Cancer TNM staging system [3].

In general, a follow-up examination was performed in our outpatient department every 3 months for the first 2 years and 6 months thereafter. The routine follow-up examination included a physical and routine blood examinations, blood chemistry, measurement of tumor markers (carcinoembryonic antigen, squamous cell carcinoma antigen), radiograph of the chest, and ultrasound. Computed tomography of the chest and upper abdomen were done every 6 months. Endoscopy was done yearly. Survival time was defined as the period from the date of surgery till death (including surgical death and non-cancer related death) or the most recent follow-up. The duration of follow-up ranged from 1 to 128 months (mean 29.8 months, median 24.0 months).

The normally distributed continuous data were described as mean ± standard deviation. Categorical data were describes as counts and proportions. Continuous variables were compared by student t test. The Pearson Chi-square test was used to compare categorical variable. The survival time was calculated by the Kaplan-Meier method, and the log rank test was used to assess the differences in survival between groups. To determine an ideal cutoff value for tumor length, the relationship between tumor length and death from esophageal cancer was investigated by using a scatter plot of the variable versus Martingale residuals from a Cox proportional hazard regression model without the variable of interest. A smoothed line fit of the scatter was then applied to detect the ideal cutoff value [16]. Based on the cutoff value, the tumor length could be treated as a categorical variable. Univariable Cox regression models were fitted to assess the relative effect of categorically and continuously coded tumor length and other clinicopathological variables on overall survival. The predictive accuracy of each clinicopathological variable was determined and was defined as the ability to discriminate between patients who died from cancer. The predictive accuracy was assessed with Harrell's concordance index (C-index) [17], which is an approximation of area under curve for time-to-event data. A C-index of 0.5 is equal to chance discrimination and a C-index of 1.0 represents a perfect discrimination. Multivariable Cox proportional hazards models were fitted to identify independent prognostic factors. A backward procedure based on the Akaike Information Criterion (AIC) was used for variable selection.

The parameters of the TNM staging system for esophageal squamous cell carcinoma (T stage, N stage, Grade and Location) were selected as a multivariable base model. Predictive accuracy of the TNM staging base model was then compared on the addition of tumor length. Multivariate regression coefficients of the predictive variables were used to develop nomogarms. Model performance was internally validated by measuring both discrimination and calibration [17]. Discrimination was evaluated by C-index as mentioned previously. Calibration was performed by a calibration curve, in which predicted versus actual survival are graphically depicted. Both discrimination and calibration were evaluated on this cohort using bootstrapping with 200 resamples [17]. To assess the clinical usefulness of prediction models, decision curve analysis was used by visualizing the net benefits of prediction models when different threshold probabilities were considered [18, 19].

For all statistical tests, two sided P < 0.05 was regarded as statistically significant. All statistical analyses were performed using SPSS version 17.0 (SPSS, Chicago, IL), and R software version 3.1.3 (https://​www.​r-project.​org/​).

Tumor length ranged from 0.3 to 23.0 cm (mean, 4.5 cm; median, 4.5 cm). The frequency distribution of tumor length for the entire cohort patients was shown in Fig. 2. Martingale residuals suggested 4 cm was an ideal cutoff value for tumor length (Fig. 3). On the basis of this cutoff value, patients were then divided into two groups (≤4 cm versus > 4 cm). Comparison of clinicopathological characteristics between these two groups was shown in Table 1. Tumor length > 4 cm significantly correlated with younger age (P = 0.023), male (P < 0.001), lower location (P = 0.01), increasing T stage (P < 0.001), worse N stage (P < 0.001), and more resected lymph nodes (P < 0.001), whereas no association with differentiation, vascular invasion, and perineural involvement could be found.

Univariable analysis identified both categorically (P < 0.001) and continuously (P < 0.001) coded tumor length were significant prognostic factors for overall survival (Table 2). The median survival time for patients with tumor length ≤ 4 cm was 48 months (95 % CI 40.8–55.2 months), whereas for those with tumor length > 4 cm it was 27 months (95 % CI 24.3–29.7 months) (P < 0.001) (Fig. 4). Other significant prognostic factors included sex (P = 0.025), differentiation (P < 0.001), T stage (P < 0.001), N stage (P < 0.001), vascular invasion (P = 0.038), and perineural involvement (P < 0.001) (Table 2). To assess predictive accuracy for each clinicopathological variable, C-index was calculated. Among all of the clinicopathological variables, tumor length was found to be the third best predictor (58.1 % as a continuous variable, 56.1 % as a categorical variable) after N stage (67.1 %) and T stage (60.5 %) (Table 2).

In Cox multivariate analysis, variable selection based on backward method using AIC was preformed. Both categorically (P = 0.018) and continuously (P < 0.001) coded tumor length were independent prognostic factors for overall survival. Other independent prognostic factors included age, differentiation, T stage, N stage, and number of resected lymph nodes. Sex, tumor location, vascular invasion and perineural involvement did not have significant impact on overall survival (Table 3).

Three prediction models were built. The first was a TNM staging base model. The second and the third were added categorically coded and continuously coded tumor length to the base model respectively. Results of three multivariate regression models were listed in Table 4. Differentiation, T stage, and N stage were independent prognostic factors in each of the three models. Both categorically and continuously coded tumor length reached statistical significance. Tumor location did not reach statistical significance in each of the three models. Three nomograms were developed for predicting overall survival based on beta coefficients in associated models (Fig. 5). Model performance was evaluated by internal validated by bootstrapping. The bootstrap-corrected C-index for TNM staging base model was 69.4 %. The addition of categorically and continuously coded tumor length to the TNM staging base model led to an increased bootstrap-corrected C-index of 69.6 and 69.8 %, respectively. The calibration curves of the three prediction models were shown in Fig. 6. Each calibration curve showed good agreement between predicted and actual outcomes. In the decision curve analysis, three models performed similarly across a wide range threshold probabilities. Models including tumor length (either categorically or continuously coded) did not show any net benefit for predicting overall survival compared to the TNM staging base model (Fig. 7).