Clinical Colorectal Cancer February 2003 • 239 Introduction Colorectal cancer is a major health concern. In the United Kingdom, there were 34,310 cases and 16,720 deaths in 1999. 1 It is the second biggest cause of cancer-related death after lung can- cer. Despite the fact that approximately 80% of patients undergo potentially curative surgery, half these patients will experience re- currence and die of their disease, most within 5 years. The role of 5-fluorouracil (5-FU)–based adjuvant chemotherapy in patients with stage III disease (T1-4 N1) is generally accepted, with an absolute survival benefit of approximately 7%. 2 The role in stage II patients (T3-4 N0) is less well defined, with most studies demonstrating a trend toward improved survival without reaching statistical significance. 3 Present prognostic systems, such as the Dukes or TNM staging classifications, subdivide the population of patients into large cat- egories with similar outcomes according to depth of bowel wall penetration and lymph node involvement. 4 However, these cate- gories are too imprecise for detailed predictions of patient out- comes. Within the category of Dukes C (stage III), 40% of pa- tients will be free of disease at 5 years, which rises to approxi- mately 47% after adjuvant chemotherapy. This means that giving adjuvant chemotherapy to all stage III patients, as is generally ac- cepted, exposes 40% of patients to chemotherapy unnecessarily, and 53% receive chemotherapy that will not prevent their disease from recurring. It is clear that present prognostic categories are too broad, and further refinement is required to more accurately predict which individual patients require further therapy. A comprehensive literature search of prognosis and colorectal neoplasms demonstrates the enormous amount of research un- dertaken to attempt to identify potential factors that might facili- tate estimation of individual patient outcomes. Extensive surveys of the literature by the American Society of Clinical Oncology (ASCO) and a recent review in the British Journal of Cancer have identified a great number of potential prognostic markers, some of which are listed in Table 1. 5,6 However, despite significant ef- forts, the ASCO expert panel believed there was insufficient evi- dence to recommend the use of any of the potential markers in everyday clinical practice, with the exception of carcinoembry- onic antigen for monitoring disease. It is useful to think of tumor markers in terms of both prognostic factors, relating to the inher- ent behavior of the malignancy, and predictive factors, relating to the response of the tumor to a particular treatment modality. In order to improve the prediction of outcome for individual patients, large volumes of demographic, clinical, and pathologic data need to be analyzed. The aim is to determine a phenotype that predicts outcome. There is controversy over the best way to analyze such data. Traditionally, multivariate analysis and para- metric regression has been used. However, the recent explosion of data mining and bioinformatics relating to huge data sets (for ex- Neural Networks in the Prediction of Survival in Patients with Colorectal Cancer Abstract It is important to predict outcome for colorectal cancer patients following surgery, as almost 50% of patients un- dergoing a potentially curative resection will experience relapse. It is clear that present prognostic categories such as Dukes or TNM staging are too broad, and further refining is required to prognosticate for high-risk subgroups. One approach is to determine a phenotype associated with recurrence.We compared 2 methods of analyzing such data. Pathologic data from a large clinical trial was analyzed for 403 patients.The outcome modeled was disease re- currence.The results from logistic regression analysis and a neural network approach are compared with respect to receiver operator characteristic plots, which estimate the fit of the model.The best logistic regression model gives a result of 66%, and the neural network approach 78%.The conclusion from this study is that the neural network ap- proach is superior to regression analysis. Further analyses are in progress using a larger patient sample size (n > 1000), improved statistical models, and a more refined neural network. Clinical Colorectal Cancer, Vol. 2, No. 4, 239-244, 2003 Key words: Dukes staging, Logistic regression, Multilayer perceptron, Prognosis, Risk factors, TNM staging Submitted: Jul 18, 2002; Revised: Sep 24, 2002; Accepted: Oct 10, 2002 Address for correspondence: Simon Grumett, MD, Institute for Cancer Studies, University of Birmingham, Birmingham B15 2TT, UK Fax: +44-121-414-3263; e-mail: simong@cancer.bham.ac.uk 1 Institute for Cancer Studies, University of Birmingham, UK 2 Xaim Inc, Colorado Springs, CO 3 Department of Clinical Pharmacology, University of Oxford, UK Simon Grumett, 1 Pete Snow, 2 David Kerr 3 Contribution Original