Colonic carcinoma is one of the most frequent malignant tumors in

Transcription

1 958 Improved Prediction of Recurrence after Curative Resection of Colon Carcinoma using Tree-Based Risk Stratification Martin Radespiel-Tröger, M.D. 1 Werner Hohenberger, M.D. 2 Bertram Reingruber, M.D. 2 1 Department of Medical Informatics, Biometry, and Epidemiology, University of Erlangen-Nuernberg, Erlangen, Germany. 2 Department of Surgery, University of Erlangen- Nuernberg, Erlangen, Germany. The authors thank Paul Hermanek, M.D., and Susanne Merkel, M.D. (both of the Department of Surgery, University of Erlangen-Nuernberg, Erlangen, Germany), for their valuable comments during data analysis and during the preparation of the article. Address for reprints: Martin Radespiel-Tröger, M.D., Department of Medical Informatics, Biometry, and Epidemiology, Friedrich-Alexander-University Erlangen-Nuernberg, Waldstrasse 6, D Erlangen, Germany; Fax: (011) ; radespmn@imbe.imed.uni-erlangen.de Received July 11, 2003; revision received November 3, 2003; accepted December 11, BACKGROUND. Patients who are at high risk of recurrence after undergoing curative (R0) resection for colon carcinoma may benefit most from adjuvant treatment and from intensive follow-up for early detection and treatment of recurrence. However, in light of new clinical evidence, there is a need for continuous improvement in the calculation of the risk of recurrence. METHODS. Six hundred forty-one patients with R0-resected colon carcinoma who underwent surgery between January 1, 1984 and December 31, 1996 were recruited from the Erlangen Registry of Colorectal Carcinoma. The study end point was time until first locoregional or distant recurrence. The factors analyzed were: age, gender, site in colon, International Union Against Cancer (UICC) pathologic tumor classification (pt), UICC pathologic lymph node classification, histologic tumor type, malignancy grade, lymphatic invasion, venous invasion, number of examined lymph nodes, number of lymph node metastases, emergency presentation, intraoperative tumor cell spillage, surgeon, and time period. The resulting prognostic tree was evaluated by means of an independent sample using a measure of predictive accuracy based on the Brier score for censored data. Predictive accuracy was compared with several proposed stage groupings. RESULTS. The prognostic tree contained the following variables: pt, the number of lymph node metastases, venous invasion, and emergency presentation. Predictive accuracy based on the validation sample was (95% confidence interval [95% CI], ) for the prognostic tree and (95% CI, ) for the UICC TNM sixth edition stage grouping. CONCLUSIONS. The prognostic tree showed superior predictive accuracy when it was validated using an independent sample. It is interpreted easily and may be applied under clinical circumstances. Provided that their classification system can be validated successfully in other centers, the authors propose using the prognostic tree as a starting point for studies of adjuvant treatment and follow-up strategies. Cancer 2004;100: American Cancer Society. KEYWORDS: colon, carcinoma, recurrence, risk, trees, classification, forecasting. Colonic carcinoma is one of the most frequent malignant tumors in industrialized countries. There always has been great interest in predictors of recurrence after resection of colon carcinoma with curative intent. If tumor recurrence after resection could be identified early enough, then potentially curative reoperations might be carried out. Moreover, adjuvant treatment could be administered to patients at high risk of recurrence, perhaps even if their lymph node status was negative. For many decades, the Dukes classification system, 1 along with its subsequent modifications, 2,3 has been the standard method of determining prognosis after surgery. In Germany, general recom American Cancer Society DOI /cncr.20065

2 Recurrence Prediction in Colon Ca/Radespiel-Tröger et al. 959 mendations have been set up based on the International Union Against Cancer tumor-lymph node-metastasis (UICC-TNM) stage grouping for colon carcinoma and are used to help select patients for adjuvant systemic treatment and to determine the extent and frequency of follow-up investigations. 4 Further prognostic factors have been identified subsequently, leading to the proposal of new prognostic classifications. 5 8 However, to our knowledge, none of the newer classifications have been accepted generally. Their broad-scale application is hampered by the fact that there is no common standard regarding the inclusion of newer prognostic factors. Furthermore, the results of prognostic factor studies based on traditional regression analysis often require the aid of computer software so that the results may be applied to future studies. In contrast, the results of the statistical method used in the current study (recursive partitioning) can be interpreted easily and applied to future studies. MATERIALS AND METHODS This study was based on data from the Erlangen Registry of Colorectal Carcinoma (ERCRC). Patients were included who fulfilled the following criteria: 1) curative resection (R0) of colon carcinoma by means of colon resection and formal lymph node dissection or, in patients with tumors located between 2 drainage areas (lateral transverse colon and hepatic or splenic flexure), extended hemicolectomy or subtotal colectomy with dissection of 2 lymphatic drainage areas; and 2) definitive tumor operation between January 1, 1984 and December 31, 1996 at the Department of Surgery of the University of Erlangen, Germany, and minimum follow-up of 2 years after surgery (end of follow-up, December 31, 1998). Furthermore, the following exclusion criteria were defined: 1) carcinoma arising in adenomatosis (n 16 patients), ulcerative colitis (n 18 patients), Crohn disease (n 4 patients), and family history of colorectal carcinoma (n 30 patients); 2) history of other neoplastic disease (n 111 patients); 3) synchronous colorectal carcinoma (n 122 patients); 4) neoadjuvant treatment with chemotherapy and/or radiation (n 14 patients); 5) adjuvant treatment with chemotherapy and/or radiation (n 112 patients); 6) presence of distant metastasis at the time of definitive surgery (n 419 patients); 7) missing values in 1 covariates (n 21 patients) except local tumor invasion beyond 15 mm of the subserosa (for explanation, see Statistical Methods, below); and 8) loss to follow-up immediately after curative resection (n 33 patients). Applying the above criteria, 641 patients were included in the study. General epidemiologic data, clinical findings, treatment, histopathologic examination, and follow-up data were collected prospectively. Detailed documentation of the histopathologic findings allowed classification of the patients in accordance with the UICC-TNM staging system. 3 The endpoint was duration of disease-free survival. (DFS) The length of DFS was measured from the date of resection until locoregional recurrence, distant metastatic recurrence, death from any cause, or December 31, 1998, whichever event occurred first. The terminal event was defined as locoregional or distant recurrence. Prognostic Factors The following prognostic factors were analyzed: age, gender, site in colon, site in left colon (i.e., descending or sigmoid), UICC pathologic tumor (pt) classification, local tumor invasion beyond 15 mm of the subserosa, pathologic lymph node (pn) classification, number of histologically examined lymph nodes, number of lymph node metastases, histologic tumor type, malignancy grade, lymphatic invasion, venous invasion, emergency presentation, intraoperative spillage of tumor cells, surgeon, and time period ( vs ). Histopathologic examination was done according to a previously reported, standardized method. 9,10 Lymphatic invasion and venous invasion were classified according to the criteria of Fielding et al. 11 The classification of histologic grade of malignancy followed World Health Organization guidelines. 12 Patients were staged according to the UICC pathologic TNM classification system (6th edition, 2002). 3 Intraoperative tumor cell spillage was defined and analyzed according to the criteria of Zirngibl et al. 13 The impact of surgeons on recurrence rates at our center was analyzed by assigning a separate code to all surgeons who performed 10 operations during the study period: Surgeons who performed 10 operations were grouped together. Statistical Methods To exclude as few patients as possible from the analysis, missing values of the covariate local tumor invasion beyond 15 mm of the subserosa (training sample: n 2 patients; 0.3%; validation sample: n 39 patients; 11.5%) were imputed using the expectation-maximization (EM) algorithm provided in the SPSS software package (version 10; SPSS Inc., Chicago, IL). Patient age is described using the mean value with standard deviation (SD). Variables without normal distribution are summarized using the median with the lower and upper quartiles (Q1 and Q3, respectively) (the number of examined lymph nodes) or with the range (the number of lymph node metastases). The median duration of follow-up is given with 95% confidence intervals (95% CIs). For analysis of DFS,

3 960 CANCER March 1, 2004 / Volume 100 / Number 5 the variables age, number of lymph node metastases, number of examined lymph nodes, and year of operation are categorized according to the sample median (number of examined lymph nodes), to a commonly used cut-off point (age), or in a way that obtained sufficiently large subgroup sizes (number of lymph node metastases). Survival curves with 3-year and 5-year DFS probability, their respective standard errors (SE) in univariate survival analysis, and 95% CIs of survival probability in graphics and in the prognostic tree were calculated according to the methods of Kaplan and Meier 14 and Greenwood, 15 respectively. DFS curves were compared using the log-rank test. 16 For the purpose of subgroup analysis in patients with tree Stage II disease, the binary covariate local tumor invasion beyond 15 mm of the subserosa (as proposed by Merkel et al. 8 ) and 5 other covariates were analyzed with respect to a possible association with DFS. The resulting P values from the respective 6 additional log-rank tests were adjusted for multiple testing according to Bonferroni (i.e., the multiplication of each P value by 6). Tree Construction The following 10 candidate covariables were used for tree construction: age, gender, site in colon, UICC-pT classification, number of lymph node metastases, histologic tumor type, malignancy grade, lymphatic invasion, venous invasion, and emergency presentation. The maximum number of covariables (c) allowable in our tree-based analysis (12 variables) was estimated as c sqrt(n), in which sqrt is the square root and n is the number of outcome events. 17 Prognostic trees were calculated according to the method of Breiman et al., 18 Segal, 19 Lausen et al., 20 and Leblanc and Crowley. 21 In this method, a prognostic tree is constructed through a recursive partitioning process that divides the study sample into smaller subgroups (the so-called tree nodes) according to whether a particular covariate (the so-called predictor) is above a selected cut-off value. The choices of the selected predictor and its corresponding cut-off value are made to maximize the similarity of individuals within tree nodes with respect to their outcome. Tree nodes that are not split further are called terminal tree nodes. To obtain a tree compatible with the UICC-TNM stage grouping, we performed a stratified tree construction as follows. In the first tree-building step, four strata were defined corresponding to UICC-TNM Stages I, II, IIIA/IIIB, and IIIC. Subsequent tree construction was confined to these four predefined strata. After initial growth of a relatively large tree, a so-called pruning (cutting-off) procedure is performed based on the split statistics of all nonterminal tree nodes 21 and on five-fold cross-validation that leads to the selection of an optimally sized FIGURE 1. The prognostic tree based on the training sample (n 641 patients). Chisq: log-rank chi-square test statistic; n: number of patients; evs: events; UICC-pT: International Union Against Cancer pathologic tumor classification; RR: relative risk; 95% CI: 95% confidence interval; 5Y-DFS: 5-year disease-free survival (%). prognostic tree. With pruning, a preliminary tree with seven terminal tree nodes was calculated. Because of the comparable recurrence risk in tree nodes III 1 and III 2, we decided to combine those two tree nodes into tree Substage III (Fig. 1). The relative risk of recurrence for tree nodes compared with the tree node that had the lowest recurrence risk and the corresponding 95% CI were computed by means of a Cox regression model. 22 Tree Validation The predictive accuracy of the obtained prognostic tree was evaluated by means of an independent validation sample using a measure of predictive accuracy based on the integrated Brier score for censored data 23 and was evaluated up to 12 years of follow-up together with the respective bootstrap confidence interval. 24 The validation sample included 338 consecutive patients with colon carcinoma who also were recruited

4 Recurrence Prediction in Colon Ca/Radespiel-Tröger et al. 961 from the ERCRC and were treated between January 1, 1978 and December 31, The inclusion and exclusion criteria were defined and applied in the same way as for recruitment of the training sample. With respect to predictive accuracy, we compared our tree with the stage groupings proposed by the UICC-TNM (sixth edition), 3 Astler and Coller, 2 Dukes, 1 Merkel et al., 8,25 and Sternberg et al. 26 Kaplan Meier DFS curves for the training sample and the validation sample are given for the three stage groupings with the highest predictive accuracy according to the validation sample. The 95% confidence interval of proportions and of median follow-up were calculated using StatsDirect software 27 (version 2.2.3). All other calculations were made using programs written by the first author in the statistical programming language R 28 (version 1.5.0) in conjunction with the software packages survival, 29 maxstat, 30 ipred, 31 bootstrap, 32 Hmisc, 33 and maptree. 34 RESULTS The mean patient age was 63.2 years (SD, 11.3 years). There were 280 female patients (44%) and 361 male patients (56%). The median follow-up with respect to recurrence was 68.5 months (95% CI, months), and the median follow-up with respect to death was 75.9 months (95% CI, months). The overall censoring rate for the event of recurrence was 79% (505 of 641 patients). After 5 years of followup, DFS probability for all patients was 79% (SE, 2%). Disease-related survival after 5 years for all patients was 83% (SE, 2%). The type of first recurrence was locoregional recurrence in 11 patients (2%), simultaneous distant metastasis and locoregional recurrence in 10 patients (2%), and distant metastasis in 115 patients (18%). The type of distant metastasis was solitary liver metastasis in 68 patients (11%), liver metastasis combined with other distant metastasis in 18 patients (3%), and other distant metastasis in 44 patients (7%). The median number of histologically examined lymph nodes was 37 lymph nodes (Q1, 25 lymph nodes; Q3, 51 lymph nodes). The median number of lymph node metastases was 0 metastases (Q1, 0 metastases; Q3, 2 metastases; range, 0 27 metastases). Patient-related, tumor-related, and treatment-related variables for the training sample and the validation sample and their correlation with 3-year and 5-year DFS are summarized in Table 1. The following factors were associated significantly with DFS in univariate analysis: pt classification (depth of local tumor invasion), local tumor invasion beyond 15 mm of the subserosa, number of lymph node metastases, pn classification, venous invasion, lymphatic invasion, histologic tumor type, malignancy grade, and emergency presentation. Conversely, age, gender, site in colon, site in left colon, number of histologically examined lymph nodes, surgeon, intraoperative tumor cell spillage, and time period were not found to be associated significantly with DFS in univariate analysis. The multifactorial, tree-based analysis of DFS lead to a prognostic tree (Fig. 1) with six risk groups based on the prognostic factors depth of invasion, number of lymph node metastases, venous invasion, and emergency presentation. In Figure 1, the relative risk of recurrence compared with Stage I in patients with Tree Substages II,II, III, III, and III is shown together with the corresponding 5-years DFS probability and respective 95% CIs. According to our tree, patients who had Stage II tumors without emergency presentation had a 2.9-fold increased risk of recurrence (95% CI, ) compared with patients who had Stage I tumors. Patients who had Stage II tumors with emergency presentation were associated with an 8.6-fold increased risk of recurrence (95% CI, ) compared with patients who had Stage I tumors. The recurrence risk for patients with absent venous invasion and with only 1 lymph node metastasis was 2.5-fold greater (95% CI, ) compared with the recurrence risk for patients with Stage I tumors. Patients without venous invasion and with 2 or 3 lymph node metastases and patients with venous invasion and 1 3 lymph node metastases had an 8.4- fold greater risk of recurrence (95% CI, ) compared with patients who had Stage I disease. In the presence of 3 lymph node metastases, the relative risk of recurrence was 16.4 times greater (95% CI, ) compared with the relative risk for patients who had Stage I tumors. Kaplan Meier DFS curves for the training sample and the validation sample according to our tree and according to the UICC-TNM and the Merkel stage groupings 8,25 are shown in Figures 2 4. Table 2 shows the results of the comparison of all prognostic classifications in terms of predictive accuracy when applied to the training and validation samples. The highest predictive accuracy was yielded by our tree (0.230; 95% CI, ), followed by the Merkel stage grouping (0.217; 95% CI, ), the UICC-TNM stage grouping (0.212; 95% CI, ), the Astler Coller stage grouping (0.168; 95% CI, ), the stage grouping according to Sternberg et al. 26 (0.145; 95% CI, ), and the Dukes stage grouping (0.143; 95% CI, ). Tables 3 and 4 describe the differences in the distribution of patients in the validation sample according to our tree and according to the UICC-TNM and Merkel stage groupings. From Table 3, it can be seen that the composition of Stage I and Stage III /IIIC are identical in the tree and in the UICC-TNM stage

5 962 CANCER March 1, 2004 / Volume 100 / Number 5 TABLE 1 Summary of Patient-Related, Tumor-Related, and Treatment-Related Variables in the Training Sample and the Validation Sample and their Correlation with Disease-Free Survival Training sample Validation sample Variable No. of patients (%) DFS; % (SE) DFS; % (SE) 3-yr 5-yr P value (log-rank) No. of patients (%) 3-yr 5-yr P value (log-rank) Total 641 (100) 81 (2) 79 (2) 338 (100) 83 (3) 80 (3) Age yrs 370 (58) 80 (3) 79 (3) 178 (53) 83 (3) 80 (4) 65 yrs 271 (42) 82 (3) 78 (3) 160 (47) 84 (4) 79 (4) Gender Female 280 (44) 80 (3) 77 (3) 161 (48) 83 (4) 79 (4) Male 361 (56) 82 (3) 80 (3) 177 (52) 83 (4) 80 (4) Site in colon Cecum/appendix 57 (9) 81 (7) 79 (7) 22 (7) 76 (12) 71 (14) Ascending colon 89 (14) 80 (6) 77 (6) 37 (11) 80 (8) 80 (8) Hepatic flexure 31 (5) 83 (8) 83 (3) 18 (5) 93 (7) 85 (11) Transverse colon 55 (9) 87 (5) 85 (6) 29 (9) 92 (6) 84 (9) Splenic flexure 27 (4) 77 (11) 73 (12) 15 (4) 86 (11) 79 (14) Descending colon 42 (7) 70 (10) 64 (12) 22 (7) 86 (9) 86 (9) Sigmoid colon 340 (53) 82 (3) 80 (3) 195 (58) 82 (3) 79 (4) Site in left colon Yes 382 (60) 81 (3) 78 (3) 217 (64) 82 (3) 79 (4) No 259 (40) 82 (3) 79 (3) 121 (36) 85 (4) 80 (5) Emergency presentation Yes 60 (9) 60 (11) 60 (11) 23 (7) 57 (2) 52 (23) No 581 (91) 83 (2) 81 (2) 315 (93) 85 (2) 81 (3) Histologic tumor type Adenocarcinoma 555 (87) 80 (2) 77 (2) 281 (83) 84 (3) 81 (3) Mucinous adenocarcinoma 86 (13) 89 (4) 86 (4) 57 (17) 80 (7) 74 (8) Malignancy grade (10) 87 (5) 85 (5) 61 (18) 93 (4) 91 (4) (74) 84 (2) 81 (2) 225 (67) 85 (3) 82 (3) 3/4 100 (16) 63 (8) 61 (8) 52 (15) 61 (12) 54 (13) pt classification pt1 59 (9) 93 (4) 90 (5) 21 (6) 100 (0) 100 (0) pt2 81 (13) 97 (2) 97 (2) 34 (10) 97 (3) 94 (4) pt3 407 (63) 80 (3) 77 (3) 239 (71) 82 (3) 78 (4) pt4 94 (15) 65 (8) 62 (8) 44 (13) 70 (10) 68 (11) Invasion 15mm of subserosa Yes 367 (57) 75 (2) 71 (2) 239 (71) 79 (3) 74 (3) No 274 (43) 90 (2) 89 (2) 99 (29) 93 (3) 92 (3) pn classification pn0 408 (64) 89 (2) 87 (2) 211 (62) 93 (2) 91 (2) pn1 160 (25) 76 (5) 73 (5) 82 (24) 78 (6) 74 (7) pn2 73 (11) 47 (13) 42 (15) 45 (13) 47 (16) 35 (21) No. of lymph node metastases (64) 89 (2) 87 (2) 211 (62) 93 (2) 91 (2) (25) 76 (5) 73 (5) 82 (24) 78 (6) 74 (7) (7) 47 (17) 47 (17) 27 (8) 59 (16) 43 (23) (4) 47 (21) 33 (29) 18 (5) 26 (42) 20 (51) Venous invasion Yes 186 (29) 65 (6) 61 (6) 83 (25) 70 (7) 64 (8) No 455 (71) 88 (2) 86 (2) 255 (75) 88 (2) 85 (3) Lymphatic invasion Yes 334 (52) 73 (3) 70 (4) 125 (37) 71 (6) 64 (7) No 307 (48) 90 (2) 88 (2) 213 (63) 90 (2) 89 (3) No. of examined lymph nodes (51) 82 (3) 80 (3) 242 (72) 79 (3) 76 (4) (49) 80 (3) 77 (3) 96 (28) 92 (3) 88 (4) Intraoperative tumor cell spillage Yes 9 (1) 78 (18) 67 (24) 10 (3) 78 (18) 78 (18) No 632 (99) 81 (2) 79 (2) 328 (97) 83 (3) 80 (3) Time period a / b 340 (53) 79 (3) 76 (3) 211 (62) 82 (3) 80 (4) a / b 301 (47) 84 (3) 82 (3) 127 (38) 85 (4) 79 (5) Surgeon A 33 (5) 78 (9) 78 (9) 25 (7) 82 (10) 77 (12) B 13 (2) 76 (16) 76 (16) 17 (5) 94 (6) 94 (6) C 38 (6) 77 (9) 77 (9) 11 (3) 71 (20) 71 (20) D 195 (30) 84 (3) 81 (4) 113 (33) 86 (4) 80 (5) E 60 (9) 82 (6) 80 (7) 24 (7) 83 (9) 79 (11) Other 302 (47) 80 (3) 77 (3) 148 (44) 80 (4) 79 (4) DFS: disease-free survival; SE: standard error; pt: pathologic tumor classification; pn: pathologic lymph node classification; log-rank: log- rank test. a Time period for training sample. b Time period for validation sample.

6 Recurrence Prediction in Colon Ca/Radespiel-Tröger et al. 963 FIGURE 2. Kaplan Meier disease-free survival curves with respective 95% confidence intervals (vertical bars) at 2-year intervals. Disease-free survival in the training sample (n 641 patients) and the validation sample (n 338 patients) was grouped according to the prognostic tree (Stages I III ). The numbers of patient at risk in each subgroup are shown. FIGURE 3. Kaplan Meier disease-free survival curves with respective 95% confidence intervals (vertical bars) at 2-year intervals. Disease free survival in the training sample (n 641 patients) and the validation sample (n 338 patients) was grouped according to the sixth edition of the International Union Against Cancer TNM edition stage grouping (Stages I IIIC). The numbers of patients at risk in each subgroup are shown. grouping. Ten patients with a comparably high recurrence risk (Tree Stage II ) were identified ( upstaged ) by the tree who were classified with Stage IIA disease according to the UICC-TNM grouping, and 19 patients were downstaged from UICC-TNM Stage IIB to Tree Stage II. Furthermore, the tree identified 32 patients with comparatively low recurrence risk (Tree Stage III ) who were classified with Stage IIIB disease according to the UICC-TNM grouping, and 5 patients were upstaged from UICC Stage IIIA to Tree Stage III. In total, 66 patients (20%) were classified differently by our tree compared with the UICC-TNM stage grouping. Furthermore, based on the (to our knowledge currently unproven) assumption that groups of patients who have a 5-year DFS rate 80% benefit from intensified follow-up (IFU), we can calculate the following results (Table 3): 16 patients who did not require IFU according to the UICC-TNM grouping would have received IFU according to our tree. Thirtytwo patients who received IFU according to the UICC- TNM grouping would not have received IFU according to our tree. Overall, follow-up would have been altered in 48 patients (14%; 95% CI, 11 18%) if the tree had been utilized. Compared with the Merkel stage grouping, 8,25 29 patients were downstaged by our tree from Merkel high-risk Stage II (Stage IIHR) into Tree Stage II (Table 4). No additional high-risk patients compared with Merkel low-risk Stage II (Stage IILR) were identified by the tree. With the exception of a single patient (who was upstaged from Merkel moderate-risk Stage III into Tree Stage III ), the pattern of shift into Stage III compared with our tree was the same for the Merkel stage grouping and for the UICC-TNM stage grouping. To evaluate a possible association between clinically relevant covariates and DFS in Tree Node II, we performed additional log-rank tests in this

7 964 CANCER March 1, 2004 / Volume 100 / Number 5 TABLE 3 Risk Group Distribution of Patients in the Validation Sample According to Prognostic Tree Substage and International Union Against Cancer Substage (n 338 patients) UICC-TNM substage Prognostic tree substage I II II III III III Row total I IIA IIB IIIA IIIB IIIC Column total UICC-TNM: International Union Against Cancer tumor-lymph node-metastases classification system (6th edition, 2002). TABLE 4 Risk Group Distribution of Patients in the Validation Sample According to Prognostic Tree Substage and Merkel stage grouping (n 338 patients) Merkel substage Prognostic tree substage I II II III III III Row total FIGURE 4. Kaplan Meier disease-free survival curves with respective 95% confidence intervals (vertical bars) at 2-year intervals. Disease free survival in the training sample (n 641 patients) and the validation sample (n 338 patients) was grouped according to Merkel stage grouping (Stages I III with low-risk [LR] and high-risk [HR] divisions). The numbers of patients at risk in each subgroup are shown. TABLE 2 Predictive Accuracy in the Training Sample and the Validation Sample According to Different Risk Classifications Stage grouping Training sample (n 641 patients) Predictive accuracy Validation sample (n 338 patients) (95% CI) Prognostic tree ( ) Merkel et al ( ) UICC-TNM (2002) ( ) Astler and Coller ( ) Sternberg et al ( ) Dukes stage ( ) 95% CI: 95% confidence interval; UICC-TNM: International Union Against Cancer tumor-lymph nodemetastases classification system (6th edition). I IILR IIHR IIILR IIIMR IIIHR Column total LR: low risk; MR: moderate risk; HR: high risk. stratum with the covariates tumor site in left colon (adjusted P 0.08), local tumor invasion beyond 15 mm of the subserosa (P 1.0), histologic tumor type (P 0.78), malignancy grade (P 1.0), lymphatic invasion (P 0.43), and venous invasion (P 1.0). After adjustments for multiple testing, none of the analyzed factors demonstrated a significant association with DFS in this stratum. Therefore, we did not introduce any additional tree splits in Tree Group II. DISCUSSION Our objective was to design and validate a prognostic tree for the prediction of recurrence risk in patients with R0-resected colon carcinoma that could be used for the effective risk-related assignment of patients to adjuvant treatment as well as the risk-adapted followup, with the potential benefit of early detection and treatment of disease recurrence. The multifactorial analysis of DFS lead to a prognostic tree that con-

8 Recurrence Prediction in Colon Ca/Radespiel-Tröger et al. 965 tained the following variables: number of lymph node metastases, UICC-pT, venous invasion, and emergency presentation. Altman and Royston 35 recommend internal validation of a prognostic classification by means of patients seen in a different time period. Consequently, we used a validation sample that comprised patients who had been treated between 1978 and Based on the validation sample, we show that our tree yields significantly greater predictive accuracy compared with all other stage groupings analyzed (Table 2). Under the assumption of a benefit of IFU in groups of patients with 5-year DFS rates 80%, we found that a substantial number of patients (n 48 patients; 14%) would have been followed differently if the tree had been utilized (Table 3). There are several possible reasons for this difference: The restriction of the current study to the analysis of colon carcinoma and the inclusion of additional prognostic factors may have increased the predictive accuracy of our classification. Furthermore, it is known that prognostic trees are suited especially to describe subgroup-specific effects of prognostic factors. Because of this specific property, in the presence of subgroup-specific effects, trees carry the potential to outperform risk classifications based on traditional regression methods. According to the validation sample, higher numbers of patients with UICC Stage III disease who had a relatively low risk of recurrence (Tree Stage III ) were identified by our tree compared with the UICC-TNM stage grouping (Table 3) and the Merkel stage grouping (Table 4). In addition, our tree showed a good separation of DFS curves between Tree Substages II and II on one hand and between Tree Substages III, III, and III on the other hand (Fig. 2). In contrast, separation of DFS curves corresponding to Stage II subgroups appears to be insufficient according to the UICC-TNM stage grouping in both the training sample and the validation sample (Fig. 3). When classified according to the Merkel stage grouping, 8,25 DFS curve separation between Substages IILR and IIHR appears to be convincing in the training sample but is less convincing in the validation sample (Fig. 4). The graphic separation of DFS curves between Stage III substages appears to be satisfactory for both the UICC-TNM and the Merkel stage groupings. However, the appearance of Kaplan Meier curves may be deceptive and is no substitute for predictive accuracy. A prognostic classification may yield low predictive accuracy, despite the seemingly good separation of the corresponding Kaplan Meier curves. Similarly, global P values cannot be interpreted as measures of predictive accuracy. 23,35 In contrast, the measure employed in this study, which was based on the agreement between predicted disease-free probability and individual observed disease-free status, is an unbiased estimator of the predictive accuracy of a prognostic classification. The prognostic impact of venous invasion in patients with colorectal carcinoma has been pointed out previously. 26,36 Several studies have shown a significant prognostic association of age and gender, respectively. 5,37 However, according to other authors, the prognostic impact of age and gender appears to be nonsignificant when controlling for other known prognostic factors Some authors have described a significant impact of surgeon s expertise 41,42 and of intraoperative tumor spillage 13 on prognosis. In the current study, age, gender, site in colon, site in left colon, number of examined lymph nodes, intraoperative tumor cell spillage, surgeon, and time period were not associated significantly with DFS. An objection concerning the design of our study may refer to the combined analysis of locoregional recurrence and distant recurrence. In view of the low frequency of locoregional recurrence in the current study (3%; n 21 patients), a separate analysis of locoregional recurrence did not appear relevant. Nevertheless, the prognostic information contained in this subgroup should be used appropriately rather than ignored, because adjuvant treatment as a possible consequence of a high recurrence risk should act on all residual tumor cells, regardless of their localization. In addition, high-risk patients may be followed by means of diagnostic imaging combined with endoscopic investigations for the detection of locoregional recurrence. Comparison of our results with the results from other studies of colon carcinoma shows a heterogeneous pattern. Multifactorial prognostic studies of DFS in patients with colon carcinoma are scarce. Two previous studies at our center 8,25 analyzed patients with UICC-TNM Stage II colon carcinoma and Stage III colorectal carcinoma, respectively. In the first study, 8 the independent prognostic factors emergency presentation, primary tumor site in the left colon, and local tumor invasion beyond 15 mm of the subserosa were identified in patients with UICC Stage II colon carcinoma. In our study, we were able to confirm the prognostic impact of the factor emergency presentation in Stage II. To compare our results further with the results of Merkel et al., 8 we analyzed six covariates (including tumor site in left colon and local tumor invasion beyond 15 mm of the subserosa) with respect to a possible association with DFS in Tree Group II. Under adjustment for multiple testing, none of the analyzed covariates showed a significant association with DFS. Therefore, we did not perform any additional splits in this subgroup. In another previous study at our center, 25 pt and pn were identified as independent predictors of survival in patients with

9 966 CANCER March 1, 2004 / Volume 100 / Number 5 UICC Stage III colorectal carcinoma. In contrast, we determined two different predictive factors (the number of lymph node metastases and venous invasion) in patients with Stage III colon carcinoma that were used to construct the prognostic tree. It should be noted that, due to differences in methods and analyzed patients, comparability of our classification with the results of Merkel et al. is limited. Sternberg et al. 26 proposed a risk classification for recurrence of colorectal carcinoma based on the factors depth of local tumor invasion, venous invasion, and lymph node involvement. In their study, the preoperative carcinoembryonic antigen (CEA) serum level was not identified as an independent prognostic factor, apparently due to the large variation of CEA measurements. Conversely, several authors have pointed out the importance of preoperative CEA serum level measurement for assessing prognosis. 7,43,44 At our center, preoperative CEA measurements have been performed routinely only since Therefore, in the current study, we did not evaluate the possible association between preoperative CEA serum level and DFS. A number of multifactorial studies of survival after curative resection of colon carcinoma have been performed. Wied et al. 45 found significant associations between survival and age, Dukes stage, malignancy grade, venous invasion, and neural invasion. Shepherd et al. 46 described Dukes stage, lymph node invasion, histologic tumor type, and character of invasive margin as independent prognostic factors for disease-related survival. However, the prognostic factors associated with survival are not necessarily the same as those associated with DFS. Furthermore, the strength of an existing association with DFS may be different compared with the association with survival. For example, age was not found to have any significant association with DFS in the current study but is known to be associated with survival. Therefore, we do not believe that prediction of recurrence on the basis of the observed survival is sufficiently precise. In several studies, combined multifactorial analyses of DFS in colon and rectal carcinoma were performed. 6,26 However, previous studies suggested that colon and rectal carcinoma differ with respect to prognostic factors. 47,48 Consequently, we would not recommend a summarized analysis of colon and rectal carcinoma. We are convinced that prognostic predictions for colon carcinoma potentially are biased if they are based on a combined analysis of colon carcinoma and rectal carcinoma. Combined multifactorial studies of survival after curative resection of colon and rectal carcinoma have been performed previously. 5,37,43,49,50 The above-mentioned limitations of combined analysis of colon and rectal carcinoma also apply here. Two limitations of the current study also should be mentioned: First, we would like to point out that our tree cannot be applied readily to individual patients, because it is designed to predict group outcomes. Second, our prognostic tree is based on characteristics observed in the population of a particular geographic region (Erlangen and its respective catchment area). We cannot rule out the possibility that the distribution of patient characteristics may be different in other reference populations. Therefore, it is necessary to perform an external validation of our tree. A multicenter validation study currently is under way. 51 External validation of our tree also is required before modified recommendations for adjuvant treatment can be made. Our prognostic tree for predicting disease recurrence in patients with R0-resected colon carcinoma was superior in terms of predictive accuracy compared with stage groupings according to the UICC-TNM, Dukes stage, Astler and Coller, Merkel et al., and Sternberg et al. when applied to an independent validation sample. Furthermore, the tree identified 32 patients (9%) in the validation sample with a comparatively low risk of recurrence (Tree Stage III ) who were classified with Stage IIIB disease according to the UICC-TNM system and 10 patients (3%) with a comparably high risk of recurrence (Tree Stage II ) who were classified with Stage IIA disease according to the UICC-TNM system. The tree is based on variables and tests that are available generally and can be applied easily under clinical circumstances. Provided our classification can be validated successfully in other centers, we propose our prognostic tree as a starting point for further clinical studies of adjuvant treatment and/or risk-adapted follow-up strategies. REFERENCES 1. Dukes CE. The classification of cancer of the rectum. J Pathol Bacteriol. 1932;35: Astler VB, Coller FA. The prognostic significance of direct extension of carcinoma of the colon and rectum. Ann Surg. 1954;139: Sobin LH, Wittekind C, editors. TNM classification of malignant tumors, 6th ed. New York: John Wiley & Sons, Hermanek P, editor. [Deutsche Krebsgesellschaft. Abridged interdisciplinary guidelines]. Munich: Zuckschwerdt, Chapuis PH, Dent OF, Fisher R, et al. A multivariate analysis of clinical and pathological variables in prognosis after resection of large bowel cancer. Br J Surg. 1985;72: Griffin MR, Bergstralh EJ, Coffey RJ, Beart RW, Melton LJ. Predictors of survival after curative resection of carcinoma of the colon and rectum. Cancer. 1987;60: Harrison LE, Guillem JG, Paty P, Cohen AM. Preoperative carcinoembryonic antigen predicts outcomes in node-negative colon cancer patients: a multivariate analysis of 572 patients. J Am Coll Surg. 1997;185: Merkel S, Wein A, Günther K, Papadopoulos T, Hohenberger W, Hermanek P. High-risk groups of patients with Stage II colon carcinoma. Cancer. 2001;92:

10 Recurrence Prediction in Colon Ca/Radespiel-Tröger et al Hermanek P. [Pathohistologic investigation of tumors]. Erlangen: Perimed, Hermanek P, Altendorf A. Classification of colorectal carcinoma with regional lymphatic metastases. Pathol Res Pract. 1981;173: Fielding LP, Arsenault PA, Chapuis PH, et al. Clinicopathological staging for colorectal cancer: an International Documentation System (IDS) and an International Comprehensive Anatomical Terminology (ICAT). J Gastroenterol Hepatol. 1991;6: Jass JR, Sobin LH. Histological typing of intestinal tumours. WHO international histological classification of tumours, 2nd edition. Berlin: Springer-Verlag, Zirngibl H, Husemann B, Hermanek P. Intraoperative spillage of tumor cells in surgery for rectal cancer. Dis Colon Rectum. 1990;33: Kaplan EL, Meier P. Nonparametric estimation from incomplete observations. J Am Stat Assoc. 1958;53: Greenwood M. The natural duration of cancer. Reports on public health and medical subjects. Volume 33. London: Her Majesty s Stationery Office, Mantel N. Evaluation of survival data and two new rank order statistics arising in its consideration. Cancer Chemother Rep. 1966;50: Altman DG. Practical statistics for medical research. London: Chapman and Hall, Breiman L, Friedman J, Olshen RA, Stone CJ. Classification and regression trees. Monterey, CA: Wadsworth, Segal MR. Regression trees for censored data. Biometrics. 1988;44: Lausen B, Sauerbrei W, Schumacher M. Classification and regression trees (CART) used for the exploration of prognostic factors measured on different scales. In: Dirschedl P, Ostermann R, editors. Computational statistics: papers collected on the occasion of the 25th Conference on Statistical Computing at Schloss Reisensburg. Heidelberg: Physica-Verlag, 1994: LeBlanc M, Crowley J. Survival trees by goodness of split. J Am Stat Assoc. 1993;88: Cox DR. Regression models and life tables (with discussion). J R Stat Soc B. 1972;34: Graf E, Schmoor C, Sauerbrei W, Schumacher M. Assessment and comparison of prognostic classification schemes for survival data. Stat Med. 1999;18: Efron B, Tibshirani RJ. An introduction to the bootstrap. New York: Chapman and Hall, Merkel S, Mansmann U, Papadopoulos T, Wittekind C, Hohenberger W, Hermanek P. The prognostic imhomogeneity of colorectal Stage III. Cancer. 2001;92: Sternberg A, Sibirsky O, Cohen D, Blumenson LE, Petrelli NJ. Validation of a new classification system for curatively resected colorectal adenocarcinoma. Cancer. 1999;86: CamCode Inc. StatsDirect statistical software. Sale, U.K.: CamCode, Ihaka R, Gentleman RR. A language for data analysis and graphics. J Comput Graph Statist. 1996;5: Therneau TM. Survival a package for survival analysis in S. Rochester [software online]. Available from URL: mayo.edu/hsr/sfunc.html [accessed December 12, 2002]. 30. Hothorn T. Maxstat maximally selected rank statistics [software online]. Available from URL: src/contrib/maxstat_0.7-6.tar.gz [accessed December 12, 2002]. 31. Peters A, Hothorn T. Ipred improved predictors [software online]. Available from URL: contrib/ipred_0.7-5.tar.gz [accessed December 12, 2002]. 32. Tibshirani R. Bootstrap functions for bootstrap, jackknife, and cross-validation [software online]. Available from URL: [accessed December 12, 2002]. 33. Harrell FE. Hmisc miscellaneous functions for R and S [software online]. Available from URL: virginia.edu/biostat/s/library/r/ [accessed December 12, 2002]. 34. White D. Maptree mapping, pruning, and graphing tree models. Available from URL: contrib/maptree_1.3-3.tar.gz [accessed December 12, 2002]. 35. Altman DG, Royston P. What do we mean by validating a prognostic model? Stat Med. 2000;19: Hobday TJ, Erlichman C. Colorectal cancer. In: Gospodarowicz MK, editor. Prognostic factors in cancer, 2nd edition. New York: John Wiley & Sons, 2002: Newland RC, Dent OF, Lyttle MN, Chapuis PH, Bokey EL. Pathologic determinants of survival associated with colorectal cancer with lymph node metastases. A multivariate analysis of 579 patients. Cancer. 1994;73: Hermanek P. Gastrointestinal carcinoma are there agerelated differences in tumor behavior? Hepatogastroenterology. 1986;33: Enblad G, Enblad P, Adami HO, Glimelius B, Krusemo U, Pahlman L. Relationship between age and survival in cancer of the colon and rectum with special reference to patients less than 40 years of age. Br J Surg. 1990;77: Heys SD, O Hanrahan TJ, Brittenden J, Eremin O. Colorectal cancer in young patients: a review of the literature. Eur J Surg Oncol. 1994;20: McArdle CS, Hole D. Impact of variability among surgeons on postoperative morbidity and mortality and ultimate survival. Br Med J. 1991;302: Hermanek P, Mansmann U, Staimmer D, Riedl S, Hermanek P. The German experience the surgeon as a prognostic factor in colon and rectal cancer surgery. Surg Oncol Clin North Am. 2000;9: Wiggers T, Arends JW, Volovics A. Regression analysis of prognostic factors in colorectal cancer after curative resections. Dis Colon Rectum. 1988;31: Compton C, Fenoglio-Preiser CM, Pettigrew N, Fielding LP. American Joint Committee on Cancer Prognostic Factors Consensus Conference: Colorectal Working Group. Cancer. 2000;88: Wied U, Nilsson T, Knudsen JB, Sprechler M, Johansen A. Postoperative survival of patients with potentially curable cancer of the colon. Dis Colon Rectum. 1985;28: Shepherd NA, Saraga EP, Love SB, Jass JR. Prognostic factors in colonic cancer. Histopathology. 1989;14: Guggenmoos-Holzmann I, Hermanek P. [Multivariate analysis of colorectal carcinoma therapy results]. In: Hermanek P, editor. [Significance of the UICC-TNM system for clinical oncology]. Munich: Zuckschwerdt, 1986: National Institutes of Health. NIH consensus conference. Adjuvant therapy for patients with colon and rectal cancer. JAMA. 1990;264: Crucitti F, Sofo L, Doglietto GB, et al. Prognostic factors in colorectal cancer: current status and new trends. J Surg Oncol. 1991;2(Suppl): Phillips RK, Hittinger R, Blesovsky L, Fry JS, Fielding LP. Large bowel cancer: surgical pathology and its relationship to survival. Br J Surg. 1984;71: Wein A, Lehnert T, Liersch T, et al. Quality controlled surgery in adjuvant treatment of colon cancer UICC III: InTACT a multicenter Phase III trial [abstract]. Ann Oncol. 2002;13:74.