Use of a Kohonen Neural Network to Characterize Respiratory Patients for Medical Intervention

Abstract

Chronic Obstructive Pulmonary Disease (COPD) is one of the leading causes of respiratory hospitalisations in adults in the USA. Prognosis correlates highly with early diagnosis, however the disease may go unnoticed in its early stages. A database of 25,000 individuals with respiratory problems was received for further investigation. The reported rate of COPD in this population was 5.8%, which is fairly low. An unsupervised neural network using the Kohonen architecture was applied to the data in order to cluster patients into groups based on risk factors for COPD. The network consisted of five output neurons. After training characteristics of the groups were examined. Three of the groups consisted of patients with a high percent of risk factors for COPD. Patients in two of those groups were correctly diagnosed as having COPD, but patients in the third group were underdiagnosed for COPD. These patients should be re-examined by a pulmonologist for possible treatment of COPD. Thus Kohonen neural networks may be a useful tool for clustering patients into groups for differential medical intervention.