Artificial neural network model accurately predicts esophageal cancer survival

Artificial neural networks (ANN) incorporating molecular markers with tumor- and patient-related information can accurately predict the outcome of esophageal carcinoma, an international team of researchers reports.

While TNM classification is the standard method for predicting outcome in esophageal cancer patients, Dr. Stephen J. Meltzer of the University of Maryland School of Medicine in Baltimore and colleagues note, this system does not include information beyond the disease itself that could influence outcome, such as comorbid conditions, and it does not incorporate emerging molecular markers.The researchers set out to use ANNs, which are built to imitate the human decision making process, to develop a more detailed prognostic model.”ANNs learn by repetitive exposure to data, estimation of output error, and subsequent feedback,” Dr. Meltzer and colleagues note in their report in the April 15th issue of CancerBased on information from 418 consecutive patients with esophageal carcinoma undergoing surgery with the intent to cure, the researchers determined which of 199 variables were best suited to predicting 1- and 5-year survival. The optimal 1-year survival model included 65 variables, with an area under the receiver operator characteristic curve of (AUR) 0.883 and 78.1% sensitivity and 84.7% specificity. The 5-year survival model included 60 variables and had an AUR of 0.884 and 80.7% sensitivity and 86.5% specificity. The models were more accurate at predicting survival than a model using only TNM information, and most were more accurate than corresponding linear discriminant analysis models.While the TNM model is much simpler to apply than an ANN model, the researchers note, they have developed a Web-based prediction engine incorporating the model that makes it much easier to use.Thirty-two variables were common to the 1- and 5-year survival model, suggesting different factors influence short and long-term survival. Dr. Meltzer and his team point out that three comorbid conditions — preoperative oral glucose tolerance, respiratory function, and liver function — had a large impact on long-term survival.The researchers conclude: “Information provided by these models may improve our ability to select appropriate and effective treatments. Furthermore, the identification of influential patient-related variables ultimately may enable behavioral or comorbid disease modification to improve survival outcomes.”(Source: Cancer 2005;103:1596-1605: Reuters Health: Oncolink: May 2005.)