Genetically engineered neural networks for predicting prostate cancer progression after radical prostatectomy.

Objectives: To use pathologic, morphometric, DNA ploidy, and clinical data to develop and test a genetically engineered neural network (GENN) for the prediction of biochemical (prostate-specific antigen [PSA]) progression after radical prostatectomy in a select group of men with clinically localized prostate cancer.

Methods: Two hundred fourteen men who underwent anatomic radical retropubic prostatectomy for clinically localized prostate cancer were selected on the basis of adequate follow-up, pathologic criteria indicating an intermediate risk of progression, and availability of archival tissue. The median age was 58.9 years (range 40 to 87). Men with Gleason score 5 to 7 and clinical Stage T1b-T2c tumors were included. Follow-up was a median of 9.5 years. Three GENNs were developed using pathologic findings (Gleason score, extraprostatic extension, surgical margin status), age, quantitative nuclear grade (QNG), and DNA ploidy. These networks were developed using three randomly selected training (n = 136) and testing (n = 35) sets. Different variable subsets were compared for the ability to maximize prediction of progression. Both standard logistic regression and Cox regression analyses were used concurrently to calculate progression risk.

Results: Biochemical (PSA) progression occurred in 84 men (40%), with a median time to progression of 48 months (range 1 to 168). GENN models were trained using inputs consisting of (a) pathologic features and patient age; (b) QNG and DNA ploidy; and (c) all variables combined. These GENN models achieved an average accuracy of 74.4%, 63.1 %, and 73.5%, respectively, for the prediction of progression in the training sets. In the testing sets, the three GENN models had an accuracy of 74.3%, 80.0%, and 78.1%, respectively.

Conclusions: The GENN models developed show promise in predicting progression in select groups of men after radical prostatectomy. Neural networks using QNG and DNA ploidy as input variables performed as well as networks using Gleason score and staging information. All GENN models were superior to logistic regression modeling and to Cox regression analysis in prediction of PSA progression. The development of models using improved input variables and imaging systems in larger, well-characterized patient groups with long-term follow-up is ongoing.