Evaluation On Prognosis And Construction Of Hazard Model In Colorectal Cancer

Colorectal cancer is one of the common malignant cancers that severely harm the public health. In the United States of America, the mortality of colorectal cancer has ranked third among all malignant cancer. In China, the mortality is also increasing, currently ranking between the 4^th and the 5^th. Up to now, the TNM staging used extensively is the gold standard for colorectal cancer prognosis. However, the patients operated on at the same TNM stage do not necessarily have the same prognosis. A perfect prognostic evaluation system can be used for guiding clinical treatments and reducing healthcare cost. A more accurately and individually prognostic evaluation model, based on histologic markers and bio-molecular markers, is anticipated.Our laboratory has possessed a previous colorectal cancer data set with 221 patients. Surgically and pathologically verified 221 colorectal cancer cases were taken from the Xiaoshan tumor registry system during 1990-2000. Patient and treatment data were collected from patient records. Those who died within one month from surgery or died of causes other than colorectal cancer were excluded from the study. Pathological data were collected by the review on sections (hematoxyline & eosin stained, H.E.). The data set also includes the score of 4 immunohistochemical predictors( P53,Ki67,Bcl2,CXCR4 ). The survival data was provided by the Xiaoshan Center for Disease Control. All follow-up ended on Dec. 31, 2002. Using the previous data set, we investigated 9 immunohistochemical predictors and added the results into the data set, including NF-kB,Sydecan-1,β-catenin,CD68,PPARγ,IGF1R,IGFBP7,IR and Thymidylate Synthase. We also updated the survival data which ended on Mar.31, 2007. Of the 221 carcinomas, 121 were male and 100 female. The median age was 59, with a range of 26-85 years. The duration of follow-up varied from 1 month to a maximum of 203 months, with the mean of 92 months. TNM stageâ… ,â…¡,â…¢andâ…£accounted for 35/221, 84/221, 80/221, and 22/221, respectively. One hundred and forty-eight were low histological grade and 73 were high histological grade. One hundred and sixty-four were tubular adenocarcinoma, 26 were papillary carcinoma, 27 were mutinous adenocarcinoma, and 4 were ring-cell carcinoma or other histological type. One-year, 3-year, 5-year and 10-year cumulative survival rate was 79%, 72%, 70% and 67%, respectively. The 221 colorectal cancers were used in training the model.To confirm the model, we investigated an addition cohort of patients as validation set. Surgically and pathologically verified 288 colorectal cancer cases were taken from the Xiaoshan tumor registry system during 2001-2006. Clinical and pathological data were collected as mentioned above. All follow-up ended on Mar. 31, 2007. The archival block of representative tumor tissues were sliced for H.E. staining and immunohistochemical staining. Among the total of 288 cases, there were 167 colon carcinoma cases and 121 rectal carcinoma cases. Of the 288 carcinomas, 153 were males and 135 were females. The mean age was 64, with a range of 24-91 years. The duration of follow-up varied from 1 month to a maximum of 75 months, with the mean of 30 months. TNM stageâ… ,â…¡,â…¢andâ…£accounted for 62/288, 81/288, 114/288, and 31/288, respectively. Two hundred and fifteen-five were low histological grade and 73 were high histological grade. 225 were tubular adenocarcinoma, 12 were papillary carcinoma, 25 were mucinous adenocarcinoma, and 26 were ring-cell carcinoma or other histological type. One-year, 3-year and 5-yearr cumulative survival rate was 84%, 76% and 75%, respectively. The 221 colorectal cancers were used in validating the model.Survival analysis was performed using SPSS 16.0 for Windows. Cumulative survival rate was calculated using life-table methods. Survival curve was drawn using the univariate survival analysis based on Kaplan-Meier methods (the significance level tested by Log-rank method). The statistically significant prognostic factors identified by univariate analysis were then analyzed using multivariate Cox proportional hazard model, and a forward stepwise method was used to bring variables into the model. According to the regression coefficients determined by the final model, the prognostic index (PI) was calculated. According to the baseline function determined by the COX proportional hazard model, 1-year, 3-year and 5-year baseline hazard function were calculated by interpolation method, and the individually hazard model to predict the 1-year, 3-year and 5-year hazard was built.MedCalc software 9.3.0.0, Frank Schoonjans was used in Receive-operationg characteristic (ROC) analysis to evaluate the accuracy of PI on predicting prognosis. Then we established the criteria in grouping PI (PI grade) according to the cut off value with the maximum Youden's index, and compared the accuracy of PI grade and TNM stage on predicting prognosis.Hierarchical cluster analyses were performed by Cluster3.0 (Stanford University), and the results were visualized by Treeview (Stanford University). Complete linkage's method was used as the cluster method, utilizing Spearman rank correlation as interval measure. Comparison of the resultant clusters was made by x² test or Fisher's exact test using SPSS 13.0 statistical software (SPSS Inc, Chicago, Illinois, USA). A significant difference was identified if the P-value was less than 0.05, and a potential significance was identified when the P-value was less than 0.1. First we performed univariate survival analysis on 17 clinico-pathological markers and 13 molecular markers in training cohort. The 17 clinico-pathological markers included age, sex, location, histological type, histological grade, the depth of infiltration, metastasis in lymph node, growth pattern, lymphovascular invasion, perineural invasion, tumor budding, tumoral-lymphocytic infiltration, peritumoral-lymphocytic infiltration, Crohn's like reaction, chemotherapy, distant metastasis, and TNM stage.The 13 molecular markers included P53, Bcl2, Ki67, NF-kB, syndecan-1, P-catenin, CD68, CXCR4, PPARγ, IGF-1R, IGFBP7, IR and thymidylate synthase. Among those markers, 18 markers were identified as the prognostic factors by univariate survival analysis, including histological type, histological grade, the depth of infiltration, metastasis in lymph node, lymphovascular invasion, perineural invasion, tumor budding, tumoral-lymphocytic infiltration, peritumoral-lymphocytic infiltration, Crohn's like reaction, chemotherapy, distant metastasis, and TNM stage, P53, the numbers of macrophages (CD68 positive cells) in the invasive margin (CD68 margin), the increased expression of CXCR4 in the invasive margin of tumor (CXCR4 margin), and IGFBP7. Growth pattern had a trend toward significance. Only those markers and age were entered into multivariate analysis. Multivariate Cox proportional hazards analysis showed that age, metastasis in lymph node, perineural invasion, tumor budding, distant metastasis, P53, and IGFBP7 were independent prognostic factors for survival.Then PI was calculated according to the regression coefficient generated by multivariate COX proportional hazard model, and analyzed the accuracy of PI in diagnosing 1-year, 3-year, and 5-year survival status. By ROC analysis, PI in diagnosing 1-year survival status achieved area under the curve (AUC) of 0.833 (95% CI: 0.776-0.880), and the cut off value of 0.735, sensitivity of 83.9%, specificity of 72.0%, P=0.0001. The cohort was dichotomized according to the cut off value of 0.735, and then analyzed by ROC curves. There was statistical significance in the group with PI > 0.735 but not in the group with PI≤0.735. The group with PI > 0.735 achieved AUC of 0.716 (95% CI: 0.569-0.835), and the cut off value of 1.643, sensitivity of 77.4%, specificity of 72.2%, P=0.0065. The criteria of PI grade (PI (1-year)) were defined as: 1: PI≤0.735; 2: 0.735≤PI≤1.643; 3: PI > 1.643. The 221 cases were grouped to 3 grade followed this criteria. By univariate survival analysis, there was significant difference between each PI grade (P < 0.0001). By ROC curves analysis, there was no significant difference on accuracy in predicting 1-year survival status between PI (1-year) and TNM stage (P=0.603).By ROC analysis, PI in diagnosing 3-year survival status achieved area under the curve (AUC) of 0.828 (95% CI: 0.772-0.876), and the cut off value of 0.344, sensitivity of 85.4%, specificity of 67.8%, P=0.0001. The cohort was dichotomized according to the cut off value of 0.344, and then analyzed by ROC curves. There was a tendency to statistical significance in the group with PI <0.344, and definite statistical significance in the group with PI> 0.344. The group with PI≤0.344 achieved AUC of 0.638 (95% CI: 0.557-0.714), and the cut off value of -0.955, sensitivity of 36.3%, specificity of 94.7%, P=0.0565. The group with PI > 0.344 achieved AUC of 0.747 (95% CI: 0.621-0.848), and the cut off value of 1.538, sensitivity of 95.7%, specificity of 55.0%, P=0.0001. The criteria of PI grade (PI (3-year)) were defined as: 1: PI≤-0.955; 2: -0.955 < PI≤0.344; 3: 0.344 < PI≤1.538; 4: PI> 1.538. The 221 cases were grouped to 4 grade followed this criteria. By univariate survival analysis, there was significant difference between each PI grade (grade 1 vs 3, 1 vs 4, 2 vs 4, 3 vs 4, P < 0.0001; 1 vs 2, P=0.0011; 2vs 3, P=0.0008). By ROC curves analysis, there was significant difference on AUC in predicting 3-year survival status between PI (3-year) and TNM stage (P=0.001). PI (3-year) had higher accuracy.The result of PI in diagnosing 5-year survival status was similar to PI in diagnosing 3-year survival status, and cut off value of PI (5-year) was equal to PI (3-year). By ROC curves analysis, there was significant difference on AUC in predicting 5-year survival status between PI (5-year) and TNM stage (P=0.003). PI (5-year) had higher accuracy.Next, the individual hazard model to predict the 1-year, 3-year and 5-year hazard was built according to the baseline function determined by the COX proportional hazard model. First, 1-year (h₁), 3-year (h₃), and 5-year (h₅) baseline hazard function were calculated by interpolation method. The results showed h₁=0.0738, h₃=0.2596, and h₅=0.2930. Second, according the formula: h_t(x)=h_(0)ttexp(PI), the 1-year, 3-year, and 5-year hazard model were established as:PI was the only variable in the model. Each patient's hazard probability at the end of 1-year, 3-year and 5-year could be predicted according to the hazard model.In order to validate the built model, we calculated the PI in the validation cohort. Because there were 2 cases with missed tumor budding, only 286 cases were analyzed in model validation. Furthermore, we grouped the patients followed by the criteria of PI grade. PI in diagnosing 1-year survival status achieved AUC of 0.944 (95% CI: 0.907-0.969), and P=0.0001. PI in diagnosing 3-year survival status achieved AUC of 0.954 (95% CI: 0.908-0.981), and P=0.0001. PI in diagnosing 5-year survival status achieved AUC of 0.923 (95% CI: 0.848-0.968), and P=0.0001. After grouped with PI (1-year), 177 were grade 1, 45 were grade 2, and 64 were grade 3. PI (1-year) was a prognostic factor (P< 0.0001) by univariate survival analysis. After grouped with PI (3-year), 39 were grade 1, 100 were grade 2, 82 were grade 3, and 65 were grade 4. PI (3-year) was a prognostic factor (P < 0.0001) by univariate survival analysis. By ROC curves analysis, there was significant difference on AUC in predicting 1-year and 3-year survival status between PI (1-year) and TNM stage (P=0.003), PI (3-year) and TNM stage (P=0.042), respectively. PI (1-year) and PI (3-year) had higher accuracy. Moreover, unsupervised hierarchical cluster analysis was often used to classify molecular markers. However, classification based on combination of molecular and pathological predictors had never been performed using hierarchical cluster analysis. For this purpose, a total of 6 pathological predictors (p) and 13 immunohistochemical predictors (m) were investigated in 221 colorectal cancers. In 2007, we identified prognostic classification based on 13m, 5m (predictors with statistical significance by univariate survival analysis), 5p (including histological type, histological grade, the depth of infiltration, metastasis in lymph node, and distant metastasis), 13m5p, and 5m5p by unsupervised hierarchical cluster analysis. By univariate survival analysis, only classification based on 5m5p (cluster5m5p) had statistical significant on prognosis. Three groups were produced: group 1 including 130 cases, group 2 including 67 cases, and group 3 including 24 cases. By univariate survival analysis, there were significant difference when group 1 vs 2 (P <0.0001), and group 1 vs 3 (P=0.0146), and there was no significant difference between group 2 and 3 (P=0.4251). In addition, there was a tendency to significant difference between the two groups produced by classification based on 5m (cluster5m) (P=0.0532). When age, cluster5m5p, cluster5m, TNM stage, histological type, and histological grade were entered into multivariate COX proportional hazard model, the results showed that cluster5m5p and TNM stage were independent prognostic factors. We defined classification based on molecular and pathological predictors as molecular-pathological classification, which is superior to that based only on molecular predictors on prognosis.Micropapillary structure is identified as tight neoplastic cell tufts which lack central fibrovascular cores and are surrounded by cleft-like spaces. Up to now, there are two reports in colorectal cancer with micropapillary component (MP), and survival analysis has never been investigated. Thirty colorectal carcinomas with a micropapillary component were identified from the series of 221 colorectal carcinomas. Of these 30 carcinomas with MP, the MP ranged from 5 to 75% of the tumor area in histological sections. TNM stageâ… ,â…¡,â…¢andâ…£accounted for 3.3% (n=1), 33.3% (n=10), 60.0% (n=18), and 3.3% (n=1), respectively. Among the 221 patients, 56 cases without MP, and 15 cases with MP died of disease by the end of follow-up. Carcinomas with MP had a worse prognosis compared with those without MP (P=0.0128). MP was an independent predictor identified by multivariate COX proportional hazard model. Furthermore, survival analysis stratified by TNM stage showed MP was a prognostic factor in TNM stageâ… -â…¡(P < 0.0001) not in TNM stageâ…¢-â…£(P=0.7223). In TNM stageâ… -â…¡, MP also was an independent prognostic factor. Carcinoma with MP compared with those without MP revealed a higher percentage of high-grade tumors (P=0.0031) and higher levels of lymphovascular invasion (P=0.0476), perineural invasion (P=0.0242), positive tumor budding (P=0.0096), positive lymph node metastasis (P=0.0281) and TNM stageâ…¢-â…£(P=0.0281). However, the results stratified by T stage indicated that the presence of MP predicted more frequent positive lymph node metastasis than the absence of MP only in T1-2 stage.Based on above results, we drew the following conclusions:1. Age, metastasis in lymph node, perineural invasion, tumor budding, distant metastasis, P53, and IGFBP7 are independent prognostic factors for survival in colorectal cancer.2. Prognostic index have robust performance in predicting 1-year, 3-year, and 5-year survival status. The PI grade based on prognostic index is more accurate than TNM stage in evaluation on prognosis.3. The hazard model based on prognostic index could be used to predict individual hazard at the end of 1-year, 3-year, and 5-year, with good repeatability.4. Classification based on pathological and immunohistochemical predictors is superior to that based only on molecular predictors on prognosis, though they both have prognostic significance in colorectal cancer.5. The presence of a micropapillary component predicts more frequent lymph node metastasis in T1-2 stage and worse prognosis in TNM stageâ… -â…¡.