Evaluation of artificial neural networks in the classification of primary oesophageal dysmotility

Abstract

OBJECTIVE: Artificial neural networks (ANNs) can rapidly analyse large data sets and exploit complex mathematical relationships between variables. We investigated the feasibility of utilizing ANNs in the recognition and objective classification of primary oesophageal motor disorders, based on stationary oesophageal manometry recordings. MATERIAL AND METHODS: One hundred swallow sequences, including 80 that were representative of various oesophageal motor disorders and 20 of normal motility, were identified from 54 patients (34 F; median age 59 years). Two different ANN techniques were trained to recognize normal and abnormal swallow sequences using mathematical features of pressure wave patterns both with (ANN(+)) and without (ANN(-)) the inclusion of standard manometric criteria. The ANNs were cross-validated and their performances were compared to the diagnoses obtained by standard visual evaluation of the manometric data. RESULTS: Interestingly, ANN(-), rather than ANN(+), programs gave the best overall performance, correctly classifying >80% of swallow sequences (achalasia 100%, nutcracker oesophagus 100%, ineffective oesophageal motility 80%, diffuse oesophageal spasm 60%, normal motility 80%). The standard deviation of the distal oesophageal pressure and propagated pressure wave activity were the most influential variables in the ANN(-) and ANN(+) programs, respectively. CONCLUSIONS: ANNs represent a potentially important tool that can be used to improve the classification and diagnosis of primary oesophageal motility disorders.