BACKGROUND: Artificial neural networks (ANN) are modelling mechanisms that are highly flexible and adaptive to solve the non-linearity inherent in the relationship between symptoms and underlying pathology. OBJECTIVES: To assess the efficacy of ANN in achieving a diagnosis of gastro-oesophageal reflux disease (GORD) using oesophagoscopy or pH-metry as a diagnostic gold standard and discriminant analysis as a statistical comparator technique in a group of patients with typical GORD symptoms and with or without GORD objective findings (e.g. a positive oesophagoscopy or a pathological oesophageal pH-metry). METHODS: The sample of 159 cases (88 men, 71 women) presenting with typical symptoms of GORD, were subdivided on the basis of endoscopy and pH-metry results into two groups: GORD patients with or without oesophagitis, group 1 (N=103), and pH and endoscopy-negative patients in whom both examinations were negative, group 2 (N=56). A total of 101 different independent variables were collected: demographic information, medical history, generic health state and lifestyle, intensity and frequency of typical and atypical symptoms based on the Italian version of the Gastroesophageal Reflux Questionnaire (Mayo Clinic). The diagnosis was used as a dependent variable. Different ANN models were assessed. RESULTS: Specific evolutionary algorithms selected 45 independent variables, concerning clinical and demographic features, as predictors of the diagnosis. The highest predictive performance was achieved by a 'back propagation' ANN, which was consistently 100% accurate in identifying the correct diagnosis compared with 78% obtained by traditional discriminant analysis. CONCLUSION: On the basis of this preliminary work, the use of ANN seems to be a promising approach for predicting diagnosis without the need for invasive diagnostic methods in patients suffering from GORD symptoms.
BACKGROUND: Artificial neural networks (ANN) are modelling mechanisms that are highly flexible and adaptive to solve the non-linearity inherent in the relationship between symptoms and underlying pathology. OBJECTIVES: To assess the efficacy of ANN in achieving a diagnosis of gastro-oesophageal reflux disease (GORD) using oesophagoscopy or pH-metry as a diagnostic gold standard and discriminant analysis as a statistical comparator technique in a group of patients with typical GORD symptoms and with or without GORD objective findings (e.g. a positive oesophagoscopy or a pathological oesophageal pH-metry). METHODS: The sample of 159 cases (88 men, 71 women) presenting with typical symptoms of GORD, were subdivided on the basis of endoscopy and pH-metry results into two groups: GORD patients with or without oesophagitis, group 1 (N=103), and pH and endoscopy-negative patients in whom both examinations were negative, group 2 (N=56). A total of 101 different independent variables were collected: demographic information, medical history, generic health state and lifestyle, intensity and frequency of typical and atypical symptoms based on the Italian version of the Gastroesophageal Reflux Questionnaire (Mayo Clinic). The diagnosis was used as a dependent variable. Different ANN models were assessed. RESULTS: Specific evolutionary algorithms selected 45 independent variables, concerning clinical and demographic features, as predictors of the diagnosis. The highest predictive performance was achieved by a 'back propagation' ANN, which was consistently 100% accurate in identifying the correct diagnosis compared with 78% obtained by traditional discriminant analysis. CONCLUSION: On the basis of this preliminary work, the use of ANN seems to be a promising approach for predicting diagnosis without the need for invasive diagnostic methods in patients suffering from GORD symptoms.
Authors: Daniel R Witt; Huijun Chen; Jason D Mielens; Kieran E McAvoy; Fan Zhang; Matthew R Hoffman; Jack J Jiang Journal: J Voice Date: 2013-12-05 Impact factor: 2.009
Authors: Jason D Mielens; Matthew R Hoffman; Michelle R Ciucci; Timothy M McCulloch; Jack J Jiang Journal: J Speech Lang Hear Res Date: 2012-01-09 Impact factor: 2.297