Abstract
Neural network has attracted interests to develop the knowledge base from the medical database. Some neural networks have been applied to the diagnostic system, but not to the epidemiological analysis. Non-linear characteristics derived from neural network seems to be more appropriate to construct disease models compared to ordinary stastical methods. This study aims to clarify the applicability of neural network to the representtation of the hypertension model. From medical database, 598 cases were chosen to the learning group, and 597 cases were chosen to the testing group, randomly. Input variables used in this study were sex, age, smoking and drinking habits, body mass index, systolic and diastolic blood pressure, total cholesterol, triglyceride, fasten plasma glucose and uric acid. These input variables were used directly and also used after transforming to the fuzzy memberships. The neural network was learned based on the back propagation of error. The diagnostic accuracies were compared among the neural network directly used input variables, the neural network used fuzzy memberships of input variables and logistic regression as the ordinary methods. The neural network used fuzzy membership showed highest diagnostic accuracy than other methods. This neural network was also able to represent the risk factors associated to the occurrences of hypertension.
