Parallel Combinations of Intelligent Systems in Classifying Tracheal-bronchial Breath Sounds Using a Novel Cross- Validation Technique

Respiratory auscultation is a fundamental tool in assessing the physiological state of a patient. Second opinions are regularly sought by medical practitioners which allows for a framework of reasoning between different hypotheses in establishing a final diagnosis. Here we assess two such reasoning mechanisms: a novel, structured approach of serially propagating knowledge from a Fuzzy C-Means (FCM) system to a Self Organising Map (SOM); and a more traditional parallel approach of multi-classifier combination using DempsterShafer Theory. We have demonstrated that in the context of breath sound classification the parallel system attained a classification accuracy of 87% (87.25% sensitivity) compared to individual classifiers such as the Multi-Layer Perceptron (72% accuracy, 72.75% sensitivity) and the Radial Basis Function Network (80% accuracy, 80.75% sensitivity). We have further demonstrated the validity of this combined approach with the FCM/SOM method of cluster analysis.