The Use of Artificial Neural Networks for the Diagnosis and Estimation of Prognosis in Cancer Patients

Abstract Artificial neural networks (ANNs) are computational models that attempt to emulate the architecture of the human brain. Data is processed in discrete “elements” or “neurons” and transferred amongst these elements through “connections” that simulate neural synapses. Artificial neurons are usually organized in the following layers: input, hidden and output. The neurons in the input layer compute the data provided by the investigators; the elements in the single or multiple hidden layers allow for the development of multiple variations of the data; while the output layer calculates the “answer” estimated by the model. Various mathematical functions are applied to the data, resulting in several ANN models, such as backpropagation, probabilistic and others. ANNs offer the advantage of being able to work with “uncertain” data and to benefit from the availability of numerous elements of information that may not appear on an intuitive basis to influence the solution of a particular problem. They are suitable models for the solution of classification and forecasting problems, as they calculate a value for individual members of a population. ANNs have been applied in our laboratory and elsewhere to estimate the prognosis of cancer patients. For example, in a study of patients with non-small-cell carcinoma of the lung, probabilistic ANN models were able to accurately forecast, based on several clinico-pathologic features, whether individuals with the disease would be dead or alive at 5 years after diagnosis and initial treatment. In a recent study, ANNs could classify cell lines with small cell carcinoma and non-small-cell carcinoma of the lung, based on DNA hypermethylation data collected with molecular methods. ANNs and other multivariate classificatory and forecasting models can be influenced by data “overfit”, resulting in estimations that are mostly due to chance. The results of various ANN models need to be validated with jackknife analysis and other methods. Potential applications of ANNs for the diagnosis and estimation of prognosis in cancer patients are discussed, and the difficult problem of model accuracy validation is reviewed.