Commentary Predicting Survival with Artificial Neural Networks Romano Danesi Antonello Di Paolo
The term neural network is used to describe a wide range of computational architectures used for several tasks including prediction, classification of data sets, and data mining. Neural networks are made up of many processing units, each of which has a small amount of local memory.1 Communication among processing units is enabled by connections that carry numerical data and link the units to one another, so that each operates on local data and inputs received via connections from other units. In this application, an artificial neural network is used to search for dependencies among the variables in a data set. The network is used to analyze the strength of connections among various sets of simple data, and in order to perform this task, the architecture must be very computationally efficient.1 The neural network architecture’s properties are fully used when performing unsupervised pattern recognition to search for previously unknown higher-order dependencies in the data set. In addition to this, a Bayesian statistical approach has been developed to highlight unexpected dependencies in the data set in order to increase the computational robustness of the procedure.2
tion of a fast-moving field that has been revitalized by the need to search within the large amount of biological data now available as a result of molecular techniques for dependencies that will better enable predictions regarding the disease outcome.3 The application of artificial neural networks to oncology has been documented in the clinical setting. Artificial neural networks have been developed for the early detection of prostate cancer in men with abnormalities in total prostate-specific antigen (PSA), and the predictive accuracy of the model system was compared with that obtained by the use of conventional statistical analysis of standard PSA parameters.4 Results provided evidence that the predictive accuracy of the artificial neural network was superior to that of conventional analysis of standard PSA parameters, and that the application of artificial models might help select patients with early cancer who need prostate biopsy. An artificial neural network has also been developed to predict the presence of cancer following debulking laparotomy and chemotherapy in patients with International Federation of Gynecology and Obstetrics stage III or IV ovarian cancer.5 The residual gross tumor or microscopic disease was determined by a second-look laparotomy and results demonstrated that the artificial neural network performed significantly better than logistic and linear regression analyses.
Neural Networks with ProstateSpecific Antigen as a Predictive Parameter In this issue of Clinical Colorectal Cancer, Grumett et al provide an easily accessible and comprehensive descrip-
Predictive Value of Neural Networks in Genotyping Perhaps one of the most promising developments of artificial intelligence tools is in the study of biological systems; indeed, recent applications in-
Division of Pharmacology and Chemotherapy Department of Oncology University of Pisa, Italy
clude the analysis of expression profiles and genomic and proteomic sequences.6 With the completion of the Human Genome Project and the identification of genes implicated in the development of cancer, the next task will be the understanding of the influence of oncogenes and tumor suppressor genes on drug susceptibility and the application of genetic profiling to the choice of pharmacologic treatment. The influence of genetic background on responses to anticancer agents is a topic of exceptional interest, particularly if we consider that the reasons for the outstanding success of chemotherapy in selected cancers (ie, testicular seminoma) and the discouraging results in others (ie, ductal pancreatic cancer) are still unclear. Pharmacogenetics, with the help of advanced bioinformatics, will play a crucial role in therapeutic decisionmaking and treatment optimization. In conclusion, further studies are required in order to expand the use of artificial neural networks, to assess their computational efficiency and stability, and to test their effectiveness and efficiency through simulation processes. 01. Bate A, Lindquist M, Edwards IR, et al. A data mining approach for early detection and analysis. Drug Saf 2002; 25:393-397. 02. Bate A, Lindquist M, Edwards IR, et al. A Bayesian neural network method for adverse drug reaction signal generation. Eur J Clin Pharmacol 1998; 54:315321. 03. Grumett S, Snow P, Kerr D. Neural networks in the prediction of survival in patients with colorectal cancer. Clin Colorectal Cancer 2003; 2:239-244. 04. Djavan B, Remzi M, Zlotta A, et al. Novel artificial neural network for early detection of prostate cancer. J Clin Oncol 2002; 20:921-929. 05. Snow PB, Brandt JM, Williams RL. Neural network analysis of the prediction of cancer recurrence following debulking laparotomy and chemotherapy in stages III and IV ovarian cancer. Mol Urol 2001; 5:171-174. 06. Almeida JS. Predictive non-linear modeling of complex data by artificial neural networks. Curr Opin Biotechnol 2002; 13:72-76.
Clinical Colorectal Cancer February 2003