Combination of Preoperative Parameters
Based on the previous discussion, it is logical that several authors attempted to combine various individual parameters to construct models for predicting stage.
Partin and associates, Johns Hopkins University School of Medicine, used clinical stage, Gleason score, and serum PSA levels to predict pathological stage in 703 patients with localized prostate cancer. This is the most commonly used model for the prediction of prostate cancer stage. All patients were assigned a clinical stage by a urologist. Logistic regression analysis with the likelihood ratio chi-square test determined that serum PSA, Gleason score, and clinical stage all were good predictors of the pathological stage. The results improved by combining these 3 variables and this provided the best separation. In this study, the clinical stage underestimated pathological stage. Approximately 50% of clinical stage T2a, T2b or T1b cases had nonorgan-confined disease. The accuracy of these nomograms was not reported in this study.
Kattan and associates evaluated the utility of original Partins nomograms in 697 patients who underwent RP. They observed that although nomograms did discriminate well between organ-confined and nonorgan-confined cancers (concordance index value = 0.758), the concordance index was 0.742, 0.750, and 0.768 for prediction of extracapsular extension, SV and lymph node spread. There was significant departure from predicted probabilities as the prediction percentages increased. The performance was especially poor for SV and lymph node involvement. If the nomogram predicted a 75% chance of lymph node metastasis, the actual incidence was only 20%. They therefore concluded that nomograms may not be totally applicable to general urologic practice until further validation and modifications are performed.
Since then, Partin and associates have combined the clinical data from 3 academic institutions and developed a multi-institutional model combining serum PSA level, clinical stage, and Gleason score to predict pathological stage for 4133 men with clinically localized prostate cancer. In the validation analyses, 72.4% of the nomograms correctly predicted the probability of a pathological stage to about 10% (organ-confined disease, 67.3%; isolated capsular penetration, 59.6%; SV involvement, 79.6%; pelvic lymph node involvement, 82.9%).
Bostwick and associates reviewed the preoperative medical records and biopsy findings from 314 patients with clinical stages T1cN0M0 to T2cN0M0. They noted that the prognostic factors for capsular perforation and SV invasion were PSA, Gleason score, and percent cancer in the biopsy specimens. Multivariate analysis identified PSA (P = .0054) and percent cancer in the biopsy (P < .0001) to be the most important independent predictors of SV invasion. The 2-variable model with PSA and Gleason score had only slightly less predictive power than the model with PSA and percent cancer, similar to the findings with capsular perforation.
Narayan and Tewari used ultrasound-guided systematic and lesion-directed biopsies, biopsy Gleason score, and preoperative serum PSA level as 3 objective and reproducible variables to provide a reliable combination in preoperative identification of risk of extraprostatic extension in patients with clinically localized prostate cancer. The log-likelihood chi-square analyses revealed that a combination of systematic biopsy-based stage, serum PSA, and biopsy Gleason score best predicted extracapsular extension, margin positivity, SV involvement, and lymph node metastases. Probability plots were then constructed to make the nomogram user-friendly.
Badalament and associates  constructed a nomogram utilizing serum PSA, sextant core biopsy data, biopsy Gleason score, and quantitative nuclear imaging parameters. Univariate logistic regression analysis demonstrated that, in decreasing order, quantitative nuclear grade, preoperative PSA, total percent tumor involvement, number of positive sextant cores, preoperative Gleason score and involvement of more than 5% of a base and/or apex biopsy were significant (P = .006) for prediction of disease organ confinement status.
We have used systematic biopsy-based staging information with serum PSA and Gleason score in 800 patients with clinically localized prostate cancer and constructed an "Artificial intelligence-based Genetic Adaptive Neural Network" for pathological staging of prostate cancer.[131,132] Using this network, the sensitivity and specificity for margin positivity was 81.3% and 75%, respectively. Another follow-up study was performed to determine if precise biopsy mapping would improve the accuracy of neural network analysis. Age, race, percent cancer in biopsy, number of positive cores, location of positive biopsies, polymorphic CAG repeat sequence of androgen receptor gene, serum PSA, prostatic acid phosphatase, and free vs complex PSA were estimated on preoperative samples. Detailed pathological staging was performed on prostate, SVs, and lymph nodes. A Genetic Adaptive Probabilistic Neural Network Model was then created to predict patients having extensive capsular penetration as an output. The predictive performance of this model was: accuracy, 79%; sensitivity, 71%; specificity, 84%; PPV, 78%; negative predictive value, 79%; and area under the curve (AUC), 0.7670.
In another study, the utility of this modality in predicting biochemical recurrence following RP was evaluated. We studied 284 patients with a clinically localized prostate undergoing pelvic lymphadenectomy and RP. Using preoperative PSA, Gleason score, systematic biopsy data, and pathological staging findings, a genetic adaptive model was trained for 204 patients and tested on 80 patients to predict biochemical failure. The results on the validation set were as follows: sensitivity, 76%; specificity, 80%; AUC, 0.81%, and overall accuracy, 79% (Tewari et al., unpublished data).
Artificial intelligence-based neural networks are new tools that are very user-friendly and are being used in several branches of medicine, including cardiology and radiology. Use of this modality in urology has just started, and several networks have been used in decision-making in infertility and in calculus disease.[135,136,137] In the management of prostate cancer, there have been encouraging results in screening and prediction of biochemical recurrence using this modality.
A neural network predicts outcome based on the integration of multiple input variables. It is constructed from a large set of data with known values of multiple input variables and outcome endpoints. The input variables are mathematically transformed into a nonlinear equation that best classifies the data set into specific outcome groups. Computers are required to carry out such intense mathematical calculations and can be used later to apply these models to future data sets. Neural networks start the computation without any bias and tend to unmask the mathematical relationship between various inputs and outcome. They do not require linearity of data and sometimes bring out subtle and interesting relationships between various input variables, which, in turn, help to better understand the problem at hand. Traditionally, neural networks perform better than standard statistical methods if data are noisy (such as medical data). Therefore, we feel this technology can be applied to improve staging of individual patients with prostate cancer.
Currently available variables to predict the outcome of prostate cancer include clinical stage, serum PSA level, and biopsy Gleason scores; however, it is becoming increasingly apparent that these are inadequate to accurately predict the outcome. A variety of other biochemical, molecular, and genomic markers of tumor progression have been identified but have not been tested in neural network models. We therefore propose to include these newer markers with routine variables in a neural network model to improve the accuracy of predicting outcomes.
© 2000 Medscape
Cite this: Evaluation of the Patient With Prostate Cancer - Medscape - Sep 14, 2000.