Non-parametric classification of esophagus motility by means of neural networks.

Automatic long-term recording of esophageal pressures by means of intraluminal transducers is used increasingly for evaluation of esophageal function. Most automatic analysis techniques are based on detection of derived parameters from the time series by means of arbitrary rule-based criterions. The aim of the present work has been to test the ability of neural networks to identify abnormal contraction patterns in patients with non-obstructive dysphagia (NOBD). Nineteen volunteers and 22 patients with NOBD underwent simultaneous recordings of four pressures in the esophagus for at least 23 hours. Data from 21 subjects were selected for training. The performances of two trained networks were subsequently verified on reference data from 20 subjects. The results show that non-parametric classification by means of neural networks has good potentials. Back propagation shows good performance with a sensitivity of 1.0 and a specificity of 0.8.

Automatic detection of esophageal pressure events. Is there an alternative to rule-based criteria?

Ambulatory long-term motility recording is used increasingly for evaluation of esophageal function. The enormous amount of motility data recorded by this method demands subsequent computer analysis. One of the most crucial steps of this analysis becomes the process of automatic selection of relevant pressure peaks at the various recording levels. Until now, this selection has been performed entirely by rule-based systems, requiring each pressure deflection to fit within predefined rigid numerical limits in order to be detected. However, due to great variations in the shapes of the pressure curves generated by muscular contractions, rule-based criteria do not always select the pressure events most relevant for further analysis. We have therefore been searching for a new concept for automatic event recognition. The present study describes a new system, based on the method of neurocomputing. A large sample of normal esophageal pressure deflections was used as a "learning set," and the performance of the trained neural networks was subsequently verified on different sets of data from normal subjects. Our trained networks detected pressure deflections with sensitivities of 0.79-0.99 and accuracies of 0.89-0.98, depending on the recording level within the esophageal lumen. The neural networks often recognized peaks that clearly represented true contractions but that had been rejected by a rule-based system. We conclude that neural networks have potentials for automatic detections of esophageal, and possibly also other kinds of gastrointestinal, pressure variations.