Of 15,731 newly diagnosed colorectal cancer patients, there were 12,860 patients who had undergone colorectal resection for colorectal cancer (Figure ). Among them, 2,126 patients (16.7%), 4,008 patients (31.5%), 4,383 patients (34.4%) and 2,227 patients (17.5%) presented with stage I, stage II, stage III and stage IV disease, respectively. We excluded 2,227 patients with stage IV disease, 116 patients with pathological report other than adenocarcinoma or unavailable of lymph node counts, and 23 patients whose survival status could not be verified. The resulting cohort, 10,494 patients, who underwent colorectal resection in thirty-two major hospitals or cancer centers were eligible to enter into this survival tree analysis. The majority of patients were older than 70 years of age (39.8%, 4,173/10,494) and comorbidity indexes were > = 3 (52.5%, 5,509/10,494) (Table ). Among these patients, male patients were slightly more than female patients (56.6%, 5,937/10,494). Rectum was found to be the more frequent site of tumor location. About 63.3% (6,644/10,494) patients had no less than twelve lymph nodes examined and 59.0% (5,816/9,855) of pathology reports proved that surgical margin was free of tumor for more than 1 cm microscopically. In terms of involvement depth and tumor size, most patients (60.2%, 6,315/10,494) had cancer penetrated through the muscularis propria of colorectal wall into the subserosa and tumors grew larger than 4 cm (61.7%, 6,479/10,494). Variants of adenocarcinoma (such as mucinous adenocarcinoma or signet ring cell adenocarcinoma, etc.) occurred in only about 4.6% (478/10,494) of resected specimens. Stage III comprised of 41.7% (4,378/10,494) patients, followed by stage II which comprised of 38.1% (4,002/10,494) patients. Nearly half patients received chemotherapy 48.7% (5,109/10,494). However, much less patient underwent radiotherapy [10.7% (1,127/10,494)]. The duration of follow-up time was 0- 84 months with a mean of 48.9 months. P value of log-rank test for all factors was less than 0.01 (except number of lymph node examined, P = 0.013) when 5-year cancer-specific survival was designated as dependent variable. Paradoxically, the hazard ratio of radiotherapy vs. no radiotherapy was 1.21.
Schema of patients'enrollment in this study. (TCDB: Taiwan Cancer Database).
Demographic and treatment of colorectal adenocarcinoma patients
We started the RPA with training sample of 5310 patients with 1,296 patients (24.4%) died in the study period. Tumor staging (TNM system) was the most important factor that yielded a segment of 3,142 patients (16.0% dead) with stage I & II disease and a segment of 2,168 patients (36.5% dead) with stage III disease (p < 0.001). The same procedure continued following this splitting algorithm (Figure ). In the left segment, patient's age appeared to be the strongest factor (p < 0.001), which yielded a subgroup of 2,465 patients with age < 76.1 years (12.5% dead) and a subgroup of 677 patients with age ≥ 76.1 years (28.7% dead) (p = 0.001). No further split was possible in the node of stage I & II disease and age ≥ 76.1 years due to minimal criteria. The node with stage I & II disease and age < 76.1 years could be split into a subgroup of 849 patients with stage I disease (6.8% dead) and a subgroup of 1,616 patients stage II disease (15.5% dead) (p < 0.001). No further split was possible in the node of stage I disease and age < 76.1 years due to minimal criteria. But we could split the node with stage II disease and age < 76.1 years into a subgroup of 153 patients who had number of lymph nodes examined < 6 (32.7% dead) and a subgroup of 1,463 patients who had number of lymph nodes examined ≥ 6 (13.7% dead), both of which were terminal nodes.
Decision tree constructed by recursive partitioning analysis (training sample).
In the right segment, comorbidity score (CCI) seemed to be the strongest factor, by which yielded a subgroup of 166 patients with CCI < 2 (18.1% dead) and a subgroup of 2002 patients with CCI ≥ 2 (38.1% alive) (p = 0.001). The node that patients with stage III and CCI < 2 was a terminal node since no further split was possible. The node that patient with stage III and CCI ≥ 2 could be further split into a subgroup of 533 patients who didn't have chemotherapy treatment (50.1% dead) and a subgroup of 1,469 patients who ever have been treated with chemotherapy treatment (33.7% dead), both of which were terminal nodes. The results of the RPA process were validated with a test sample of 5,184 patients with colorectal cancer which were independent of the model building training sample (Figure ). Both results were closely correlated.
Figure 3 Decision tree constructed by recursive partitioning analysis (test sample). The plots of recursive partitioning analysis (Figure 2 and Figure 3) were obtained from 10,494 patients who were documented to have adenocarcinoma from colon and rectum (anus (more ...)
Thus five prognostic factors were identified (namely, TNM staging, age, comorbidity, number of lymph nodes examined, chemotherapy) for cancer-specific survival, resulting in seven terminal nodes. Based on mean survival time of the terminal nodes, we were able to categorized four risk groups (Table ). Group 1 (mild risk) consisted of 1,698 patients who had stage I colorectal cancer and age < 76.1 years (119 deaths in the study period). Group 2 (moderate risk) consisted of 3,129 patients who had stage II colorectal cancer, age < 76.1 years and number of lymph nodes examined ≥ 6, or stage III colorectal cancer with CCI < 2 (449 deaths in the study period). Group 3 (high risk) consisted of 4,605 patients who had stage I&II colorectal cancer and age ≥ 76.1 years or stage II colorectal cancer, age < 76.1 years and number of lymph nodes examination < 6, or stage III colorectal cancer CCI ≥ 2 with chemotherapy (1,502 deaths in the study period). Group 4 (very high risk) consisted of 1,062 patients who had stage III colorectal cancer, CCI ≥ 2 and without chemotherapy (525 deaths in the study period).
Assignment of Recursive Partitioning Analysis (RPA) Groups
Cancer-specific survival analysis using Kaplan-Meier plot and log-rank test revealed significant differences among groups (p < 0.0001, Figure ). In addition, we also utilized this RPA grouping classification to test the effects on predicting overall and progression-free survivals (3-year and 5-year respectively). Results showed good discriminating capability for this grouping classification to easily predict each outcome for all endpoints (Table ).
Survival analysis (cancer-specific survival as outcome) with Kaplan-Meier plot shows significant difference between groups (p < 0.0001). (No.: number).