A NEURAL NETWORK PREDICTS PROGRESSION FOR MEN WITH GLEASON SCORE 3 4 VERSUS 4 3 TUMORS AFTER RADICAL PROSTATECTOMY

Transcription

1 ADULT UROLOGY CME ARTICLE A NEURAL NETWORK PREDICTS PROGRESSION FOR MEN WITH GLEASON SCORE 3 4 VERSUS 4 3 TUMORS AFTER RADICAL PROSTATECTOMY MISOP HAN, PETER B. SNOW, JONATHAN I. EPSTEIN, THERESA Y. CHAN, KERRIE A. JONES, PATRICK C. WALSH, AND ALAN W. PARTIN ABSTRACT Objectives. To determine the significance of Gleason scores 3 4 (GS3 4) versus 4 3 (GS4 3) with respect to biochemical recurrence in a retrospective review of a series of men with clinically localized prostate cancer who underwent radical retropubic prostatectomy (RRP) and to develop and test an artificial neural network (ANN) to predict the biochemical recurrence after surgery for this group of men using the pathologic and clinical data. Methods. From 1982 to 1998, 600 men had pathologic Gleason score 7 disease without lymph node or seminal vesicle involvement. We analyzed the freedom from biochemical (prostate-specific antigen) progression after RRP on 564 of these men on the basis of their GS3 4 versus GS4 3 (Gleason 7) status. The Cox proportional hazards model was used to determine the importance of Gleason 7 status as an independent predictor of progression. In addition, an ANN was developed using randomly selected training and validation sets for predicting biochemical recurrence at 3 or 5 years. Different input variable subsets, with or without Gleason 7 status, were compared for the ability of the ANN to maximize the prediction of progression. Standard logistic regression was used concurrently on the same random patient population sets to calculate progression risk. Results. A significant recurrence-free survival advantage was found in men who underwent RRP for GS3 4 compared with those with GS4 3 disease (P ). The ANN, logistic regression, and proportion hazard models demonstrated the importance of Gleason 7 status in predicting patient outcome. The ANN was better than logistic regression in predicting patient outcome, in terms of prostate-specific antigen progression, at 3 and 5 years. Conclusions. A simple modification of the Gleason scoring system for men with Gleason 7 disease revealed a difference in the patient outcome after RRP. ANN models can be developed and used to better predict patient outcome when pathologic and clinical features are known. UROLOGY 56: , , Elsevier Science Inc. The most important goal of radical prostatectomy is cancer control. Although many improvements in surgical technique were made during the past several decades that permitted better cancer control and reduced perioperative morbidity, the overall recurrence rate after surgery for From the James Buchanan Brady Urological Institute and Department of Urology, Johns Hopkins Medical Institutions, Baltimore, Maryland; Xaim Incorporated, Colorado Springs, Colorado; and Department of Pathology, Johns Hopkins Medical Institutions, Baltimore, Maryland Reprint requests: Alan W. Partin, M.D., Ph.D., Department of Urology, Marburg 205A, Johns Hopkins Hospital, 600 North Wolfe Street, Baltimore, MD Submitted: June 2, 2000, accepted (with revisions): July 19, 2000 clinically localized prostate cancer ranges from 12% to 20% in several large series. 1 4 Many clinical and pathologic variables, used individually or in combination, help predict recurrence after radical retropubic prostatectomy (RRP). 5,6 One of those variables is the Gleason grading system for prostate adenocarcinoma. In his article in 1977, Gleason 7 simplified the diverse histologic differentiation patterns of tumor into a grading scale with five grades. The final Gleason score is formed by the sum of the two most prevalent grades in each specimen. For example, Gleason score 7 prostate cancer is mostly composed of tumor with Gleason grade 3 and 4 patterns. Most Gleason score 7 tumors are composed 2000, ELSEVIER SCIENCE INC /00/$ ALL RIGHTS RESERVED PII S (00)

2 of Gleason score 3 4 (GS3 4; Gleason grade 3 more prevalent than Gleason grade 4) disease. However, a substantial portion of Gleason 7 tumors have a Gleason score 4 3 (GS4 3) pattern. Few studies have investigated the difference in the outcome of the patients with GS3 4 versus GS4 3 tumors in a large cohort of men who have undergone RRP. Logistic regression (LR) models have been widely used to analyze the relationship between multiple predictor variables and a dichotomous outcome variable. However, LR models are limited because of their inadequate capacity to handle the variability and complexity of the data. Unlike LR, artificial neural network (ANN) models can learn the nonlinear patterns between the predictor variables and outcomes by simulating human decision-making using adaptation. 8 We retrospectively reviewed a large series of men with clinically localized prostate cancer who underwent surgery by a single urologist. We tried to determine the clinical significance of Gleason 7 status (GS3 4 versus GS4 3) at RRP on longterm patient outcome. We developed a Cox proportional hazards model to illustrate the importance of Gleason 7 status as an independent predictor of biochemical recurrence. In addition, we used pathologic and clinical data, including Gleason 7 status, to develop and test an ANN to predict biochemical recurrence. Finally, we compared the ability of this ANN and an LR model to predict the biochemical recurrence after RRP and demonstrated a better performance for the ANN. MATERIAL AND METHODS Between April 1982 and June 1998, 2293 consecutive men with clinically localized prostate adenocarcinoma underwent staging pelvic lymphadenectomy and anatomic RRP. The tumors were staged using digital rectal examination and routine serum prostate-specific antigen (PSA) determination. The postoperative follow-up of these men consisted of digital rectal examinations and serum PSA measurements (Hybritech Tandem-R and E [Beckman Coulter, San Diego, Calif] and the TOSOH [Tosoh Medics, South San Francisco, Calif] PSA assays) at quarterly intervals for the first year, semiannually for the second year, and annually thereafter. Bone scans were performed at the time of biochemical recurrence and on a yearly basis thereafter unless performed earlier for symptoms suggestive of distant metastasis. No patients included in this series received neoadjuvant therapy, adjuvant hormonal therapy before metastatic disease development, or immediate adjuvant radiation therapy on the basis of pathologic features. Six hundred men had Gleason 7 disease without evidence of lymph node or seminal vesicle involvement. Pathologic examination of the biopsy and surgical specimen categorized the disease of 564 men to either the GS3 4orGS4 3 subgroups according to the predominance of the Gleason grade of the specimen; 451 men (80%) had GS3 4 disease and 113 men (20%) had GS4 3 disease. The disease characteristics of these 564 men are given in Table I. The average age of the 564 men at the time of RRP was years (range 39 to 74). The mean follow-up was 5.9 years (range 1 to 17). KAPLAN-MEIER SURVIVAL ANALYSIS Actuarial survival analysis was performed using the Kaplan- Meier estimate method on men with GS3 4 versus GS4 3 disease to compare the freedom from biochemical recurrence after RRP (PSA greater than 0.2 ng/ml). The generalized Wilcoxon test was used to determine the statistical differences between the actuarial curves. All statistical analyses were performed using the Intercooled Stata 6.0 statistics and graphics data management system (Stata Corporation, College Station, Tex). COX PROPORTIONAL HAZARDS MODEL The pathologic variables were from the prostatectomy specimen (Gleason 7 status, surgical margin status, organ confinement status, established extraprostatic extension, and focal extraprostatic extension) and from the preoperative prostate biopsy specimen (Gleason score). The clinical variables consisted of the preoperative serum PSA level, clinical stage, age, and race. Univariate Cox regression analysis was used to fit the dependence of biochemical failure on all predictors individually. The Cox proportional hazards model was then used to identify the predictors with the most significant independent influence on biochemical recurrence-free survival. Forward and backward stepwise models for P 0.01 were performed. ANN AND LR MODELS There were 452 and 322 men with an available recurrence status at 3 and 5 years after RRP, respectively. One half of the men were randomly selected for training, and the remainder were used for testing (20%) and validation (30%). The input variables included the previously mentioned pathologic and clinical parameters. The output variable was either biochemical recurrence or nonrecurrence. Two different training, testing, and validation sets were used for the 3 and 5-year follow-up analyses. The same validation sets were used to compare the ANN and LR models. The ANN chosen in this application was a multilayer perceptron (MLP) 9 known for its stability and tendency not to overfit. MLPs have standard feed forward topology and successive layers of adaptive weights. 10 They typically contain one or two hidden layers of neurons, which are not treated as output units. 11 The training of the MLP was performed by a conjugate gradient descent method. 12 A simulated annealing method was used to find an MLP architecture that was close to optimum in terms of predictive accuracy and that penalized large networks to minimize overfitting. 13 This involved the examination of more than 1000 possible architectures. RESULTS OVERALL PROGRESSION Of the 564 men with Gleason score 7 disease on the RRP specimen and no lymph node or seminal vesicle involvement, 130 (23.0%) experienced biochemical recurrence after RRP. Eighty-nine men have developed biochemical evidence of recurrence without any other evidence of disease. Thirteen patients have had local recurrence without evidence of distant metastasis. Twenty-eight patients have developed distant metastases with or without evidence of local recurrence. Overall actuarial 5- and 10-year freedom from biochemical progression for this cohort of 564 men with Gleason 7 disease was 78% and 57%, respectively. UROLOGY 56 (6),

3 TABLE I. Clinical stage, preoperative Gleason score pattern and PSA level, and pathologic stage in 564 men who underwent anatomic radical retropubic prostatectomy for Gleason 7 disease and pathologic staging of GS3 4 vs. GS4 3 Characteristic (n 564) Total n (%) GS3 4 n (%) GS4 3 n (%) 1992 AJCC TNM staging classification T1a 4 (0.7) 4 (100) 0 (0) T1b 11 (2.0) 8 (72.7) 3 (27.3) T1c 177 (31.4) 149 (84.2) 28 (15.8) T2a 227 (40.2) 181 (79.7) 46 (20.3) T2b 110 (19.5) 81 (73.6) 29 (26.4) T2c 23 (4.1) 18 (78.3) 5 (21.7) T3a 12 (2.1) 10 (83.3) 2 (16.7) Serum PSA level (ng/ml) (18.8) 79 (80.6) 17 (17.3) (54.8) 233 (83.2) 47 (16.8) (21.1) 79 (73.1) 29 (26.9) (5.3) 21 (77.8) 6 (22.2) Preoperative prostate needle biopsy Gleason score pattern 5 55 (10.3) 45 (81.8) 10 (18.2) (52.4) 244 (87.1) 36 (12.9) 7(3 4) 152 (28.5) 114 (75.0) 38 (25.0) 7(4 3) 33 (6.2) 20 (60.6) 13 (39.4) 8 14 (2.6) 5 (35.7) 9 (64.3) Pathologic stage Organ confined 200 (35.5) 162 (81) 38 (19) Extraprostatic extension with 286 (50.7) 228 (79.7) 58 (20.3) negative surgical margins Extraprostatic extension with positive surgical margins 78 (13.8) 61 (78.2) 17 (21.8) KEY: PSA prostate-specific antigen; GS Gleason score. COMPARISON OF GLEASON SCORE PATTERNS OF PROSTATE NEEDLE BIOPSY AND RRP SPECIMENS Of the 564 men in this study, 534 had an available Gleason score pattern from the preoperative needle biopsy specimen. Overall, undergrading of the biopsy Gleason score pattern occurred compared with the corresponding surgical specimen Gleason score pattern in our study population. The results are summarized in Table I. GLEASON SCORE 7STATUS (GS3 4 VERSUS GS4 3) AND PROGRESSION The patients with GS4 3 disease had worse prognosis than did the patients with GS3 4 disease in terms of biochemical recurrence-free survival (P ). Figure 1 illustrates the Kaplan- Meier survival analysis of the likelihood of having an undetectable PSA level after RRP on the basis of the Gleason 7 status. COX PROPORTIONAL HAZARDS MODEL The univariate Cox regression model identified the surgical margin status as the most significant predictor of biochemical recurrence (data not shown). The Cox proportional hazards model demonstrated that surgical margin status, Gleason 7 status, and organ-confinement status are important, independent predictors of biochemical recurrence in both the forward and backward stepwise approaches using stringent exclusion criteria (P 0.01). The results of the Cox proportional hazards model are shown in Table II. COMPARISON BETWEEN LR AND ANN MODELS Overall, the ANN outperformed LR modeling in predicting 3 and 5-year biochemical recurrence after RRP. The ANN was developed with and without Gleason 7 status to confirm the additional value of Gleason 7 status in predicting biochemical recurrence after RRP. When the ANN model was used without Gleason 7 status as an input variable, the predictability of the model decreased, as evidenced by the decreases in the area under the receiver operating characteristic curve and in overall accuracy. The results of these modifications are shown in Table III. 996 UROLOGY 56 (6), 2000

4 FIGURE 1. Kaplan-Meier actuarial likelihood of PSA recurrence by Gleason 7 status (GS3 4 versus GS4 3) for 564 men with Gleason score 7 disease without lymph node or seminal vesicle involvement. The actuarial 3-, 5-, and 10-year biochemical recurrence-free rate after RRP for the 451 men with GS3 4 disease was 90%, 82%, and 60%, respectively. The same rate for the 113 men with GS4 3 disease was 72%, 61%, and 43%, respectively. TABLE II. Cox proportional hazards model for 564 men who underwent anatomic radical retropubic prostatectomy for Gleason 7 disease and pathologic staging of GS3 4 vs. GS4 3 (P <0.01 model by forward and backward stepwise procedures) Predictor Hazard Ratio Standard Error z Score P Value SM GS OC KEY: SM surgical margin status; GS7 Gleason 7 status (GS3 4 vs. GS4 3); OC organ confinement status. COMMENT For the past two decades, the Gleason system has helped urologists and pathologists to accurately grade prostate cancer. It has also been helpful, independently or in combination with other parameters, in choosing the appropriate therapeutic options and predicting the outcome after definitive therapy for prostate cancer Additionally, a significant difference in progression between patients with Gleason score 7 tumors and those with Gleason score 8 to 10 disease has been reported. 5,17,18 Recently, Stamey et al. 6 reported that the percentage of Gleason grade 4/5, in addition to cancer volume, positive lymph node findings, and intraprostatic vascular invasion, was independently associated with prostate cancer progression. However, assessing the distinction of GS3 4 versus GS4 3 disease is simpler, more reproducible, and more likely to be performed at routine pathologic examinations than accurate measurement of the percentage of Gleason grade patterns among expert pathologists. With a simple modification of subdividing Gleason score 7 diseases into GS3 4 and GS4 3 tumors, we have demonstrated that a statistically significant difference occurs in recurrence-free survival for the group of patients with Gleason 7 disease after RRP. When predicting tumor recurrence after RRP, statistical methods or the clinical judgment of experts have been used. LR is a widely used statistical modeling technique that describes the relationship between a dichotomous response (output) variable and a set of explanatory (input) variables. ANN models have emerged as an alternative to LR, since the ANN can detect complex nonlinear relationships between independent and dependent variables. 19 ANNs grew out of attempts to mimic the faulttolerance and learning capacity of biologic nervous UROLOGY 56 (6),

5 TABLE III. Comparison of artificial neural network and logistic regression models 3-Year Recurrence 5-Year Recurrence ANN ANN* LR ANN ANN* LR ROC area-validation cases U Sensitivity at 95% specificity Sensitivity at 90% specificity NPV at 95% specificity PPV at 95% specificity OA at 95% specificity NPV at 90% specificity PPV at 90% specificity OA at 90% specificity Most important variables SM, GS7 SM OC, SM, GS7 SM KEY: ANN artificial neural network; LR logistic regression, using all input variables; ROC receiver operating characteristic; U unreliable, not enough data to generate ROC area; NPV nonrecurrence predictive value; PPV recurrence predictive value; OA overall accuracy; GS7 Gleason 7 status (GS3 4 vs. GS4 3); SM surgical margin status; OC organ-confinement status. Same validation set (30% of total) was used for all ANNs and LR calculations. * ANN without GS7 status as input. systems. Like a biologic nervous system with numerous interconnected neuron cells, a typical ANN consists of computational neurons or processing elements connected together by weighted signal pathways. Properly trained ANNs have repeatedly demonstrated superior predictive accuracy to other predictive technologies, including LR, particularly when many variables are related in nonlinear ways. In addition, specific ANNs are capable of identifying variables or combinations of variables that are important to a given outcome and then ranking the order of these variables in terms of importance. In our study, one half of the men were randomly selected for training, and the remainder were used to test and validate the ANN. Overall, ANNs outperformed LR models in predicting recurrencefree outcome for our patient population. In addition to the Kaplan-Meier analysis and the Cox proportional hazards model, the ANN model confirmed that Gleason 7 status can and should serve as an important outcome predictor for patients with Gleason score 7 disease after RRP. The use of an ANN in the surgical outcome analysis for prostate cancer is a relatively new concept. In 1994, Snow et al. 20 first used ANNs to predict the recurrence of prostate cancer in 938 men after RRP. The input variables included preoperative parameters and the output variable was cancer recurrence status after RRP. Their ANN model achieved an overall accuracy of 90% in predicting cancer recurrence; however, the length of follow-up was not specified. As a result, even though their ANN was designed to predict eventual recurrence outcome, not all eventual recurrence status was available at the time of the ANN training and testing, limiting the ability of the ANN to predict long-term outcome. 8,20 In our models, we only used the parameters of the patients with known 3 or 5-year biochemical recurrence outcomes for the training, testing, and validation of the ANN. We expect an improved performance of the ANN in predicting eventual outcome after RRP as more long-term follow-up data accumulate. Finally, the general applicability of the ANN will require interinstitutional validation. CONCLUSIONS The results of this study have demonstrated the importance of Gleason 7 status in predicting patient outcome after surgery for clinically localized prostate cancer. The Kaplan-Meier survival analysis showed that patients with GS3 4 disease have a statistically significant recurrence-free survival advantage compared with those with GS4 3 disease after RRP. The Cox proportional hazards model demonstrated that Gleason 7 status is an important, independent predictor of postoperative biochemical recurrence. In addition, the ANN was better than LR in predicting progression-free survival after surgery. Finally, the ANN also confirmed the importance of Gleason 7 status, in addition to other clinical and pathologic variables, in predicting the outcome of patients with Gleason 7 disease. REFERENCES 1. Catalona WJ, and Smith DS: 5-year tumor recurrence rates after anatomical radical retropubic prostatectomy for prostate cancer. J Urol 152: , Catalona WJ, and Smith DS: Cancer recurrence and survival rates after anatomic radical retropubic prostatectomy for prostate cancer: intermediate-term results. J Urol 160: , UROLOGY 56 (6), 2000

6 3. Trapasso JG, dekernion JB, Smith RB, et al: The incidence and significance of detectable levels of serum prostate specific antigen after radical prostatectomy. J Urol 152: , Pound CR, Partin AW, Eisenberger MA, et al: Natural history of progression after PSA elevation following radical prostatectomy. JAMA 281: , Pound CR, Partin AW, Epstein JI, et al: Prostate-specific antigen after anatomic radical retropubic prostatectomy: patterns of recurrence and cancer control. Urol Clin North Am 24: , Stamey TA, McNeal JE, Yemoto CM, et al: Biological determinants of cancer progression in men with prostate cancer. JAMA 281: , Gleason D: Histologic grading and clinical staging of carcinoma of the prostate, in Tannenbaum M (Ed): Urologic Pathology. Philadelphia, Lea & Febiger, 1977, pp Kattan MW, Cowen ME, and Miles BJ: Computer modeling in urology. Urology 47: 14 21, Veelenturf LPJ: The Continuous Multi-Layer Perceptron in Analysis and Application of Artificial Neural Networks. London, Prentice Hall, 1995, pp Rummelhard DE, and McClelland JL: Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Cambridge, MIT Press, 1986, pp Bishop CM: Neural Networks for Pattern Recognition. New York, Oxford University Press, 1995, pp Bishop CM: Neural Networks for Pattern Recognition. New York, Oxford University Press, 1995, pp Kirkpatrick S, Gelatt CD, and Vecchi MP: Optimization by simulated annealing. Science 220: , Partin AW, Kattan MW, Subong EN, et al: Combination of prostate-specific antigen, clinical stage, and Gleason score to predict pathological stage of localized prostate cancer. A multi-institutional update. JAMA 277: , Bostwick DG, Qian J, Bergstralh E, et al: Prediction of capsular perforation and seminal vesicle invasion in prostate cancer. J Urol 155: , Rogers E, Gurpinar T, Dillioglugil O, et al: The role of digital rectal examination, biopsy Gleason sum and prostatespecific antigen in selecting patients who require pelvic lymph node dissections for prostate cancer. Br J Urol 78: , Epstein JI, Pizov G, and Walsh PC: Correlation of pathologic findings with progression after radical retropubic prostatectomy. Cancer 71: , Epstein JI, Partin AW, Sauvageot J, et al: Prediction of progression following radical prostatectomy: a multivariate analysis of 721 men with long-term follow-up. Am J Surg Pathol 20: , Tu JV: Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcome. J Clin Epidemiol 49: , Snow PB, Smith DS, and Catalona WJ: Artificial neural networks in the diagnosis and prognosis of prostate cancer: a pilot study. J Urol 152: , UROLOGY 56 (6),