Applying artificial neural networks to the diagnosis of organic dyspepsia

Dyspepsia diagnoses and treatment decisions are made in situations in which multiple factors must be taken into account. Evolving from neuro-biological insights, artificial neural networks (ANNs) can employ multiple factors in resolving medical prediction, classification, pattern recognition, and pattern completion. The objective of this study was to compare predictive results classifying people with organic dyspepsia with Helicobacter pylori testing (rapid urease test), a scoring system based on patients' symptoms (derived using logistic regression), classification and regression trees (CART) and the most common ANN approach used in medicine: a feed-forward multilayer perceptron (MLP) trained by back-propagation.

A scoring system, CART algorithm, and MLP model were constructed. Predictive accuracy was calculated for them and for Helicobacter pylori testing.

MLP model had a sensitivity of 0.91 (0.81 for all data) and a specificity of 0.74 (0.79 for all data) for test data. That compares favorably with Helicobacter pylori testing (sensitivity = 0.80, specificity = 0.43), the scoring system (sensitivity = 0.85, specificity = 0.60), and the CART model (sensitivity = 0.88, specificity = 0.53). Diagnostic accuracy, the area under the curve, was 0.82 using the MLP model, 0.61 using Helicobacter pylori testing, 0.78 using the scoring system, and 0.72 for the test set using CART.

The results of the analysis showed that the ANN model derived has better predictive accuracy than Helicobacter pylori testing, than a scoring system based on patients' symptoms and than a decision tree algorithm (CART). ANN model could be used as a predictive tool for organic dyspepsia and would be useful in the process of referral of dyspeptic patients from primary care to endoscopy units.