Abstract
For diagnosis and treatment of chronic hepatitis C, there have been studies to predict the stage of liver fibrosis using the results of clinical blood and urine tests. Although many of conventional studies used the present test results, given the mechanism of liver fibrosis progress via the time-varying inflammation, it is considered to be effective to use the time series of test results from past to present. We propose a combination of the mean, standard deviation, and linear predictive coding cepstrum of the time series of clinical blood and urine test results as a feature. We then examine the effectiveness of the proposed feature through experiments to evaluate the prediction performance of the proposed feature and various competitive features. We used clinical dataset on hepatitis that was publicly provided at international conferences as inputs. The nearest neighbor method was adopted as a classifier. For each feature, we estimated the prediction performance by the leave-one-out cross-validation in two-class problems, one on the seriousness of fibrosis and the other on the existence of cirrhosis. The proposed feature consistently outperformed the conventional and other competitive features for different kinds of clinical tests. Especially under the conditions using the combination of clinical tests, the proposed feature achieved higher prediction performances at 5.32 to 13.83% than did the conventional features. The results supported the highest effectiveness of the proposed feature.
