Forcasting the Clinical Outcome: Artificial Neural Networks or Multivariate Statistical Models?

Abstract

The field of prognostics has grown rapidly in the last decade and clinicians have been provided with numerous tools to assist with evidence-based medical decision-making. Most of these included nomograms, classification and regression tree analyses, and risk group stratification models [Grossberg JA, et al., 2006] and Artificial Neural Networks (ANN) [Djavan B, et al., 2002]. Nomograms are graphic representation of statistical model, which incorporate multiple continuous variables to predict a patient’s risk of developing a specific endpoint (recurrence, survival, complications) [Kattan MW, 2005]. Each variable is assigned a scale of points according to its prognostic significance. The total score for all the variables is converted to an estimated probability of reaching the endpoint [Akl A, et al., 2008]. Statistical approaches require guesses as to how outputs functionally depend on inputs. Artificial neural networks have been used for evaluation of clinical data to provide results similar to conventional modeling methods [Freeman RV, et al., 2002]. They do not require the articulation of such a mathematical model. ANNs are complex computational systems that can provide a nonlinear approach for data analysis. The ANNs forms a mapping from input to output nodes (simulated neurons) by extracting features from input patterns, assigning them weights, summing weights with activation functions, and propagating decisions to output nodes if activation thresholds are exceeded. Typical networks are organized into three layers of computational units (nodes) in which input/output layers are linked by hidden layers of nodes. Subject factors determine the number of input units, and the classification complexity determines the number of output units. The number of hidden units is determined by trial and error (training). Common routines start with one hidden unit and assign small arbitrary weights to all nodal connections. The network is fed sample data with known outcomes, and an error term is calculated by means of differences between known and predicted outputs. Learning consists of adjusting weights by backward pass of errors through the connections to network nodes in response to input data. Hidden units are added to achieve minimum error criteria, while constraining the number to promote generalization of input patterns and prevent overfitting (memorization). Interconnection density determines the network’s ability to correctly discriminate the outcomes. In statistical parlance, ANN models are a form of nonlinear