頁籤選單縮合
題名 | Wavelet Networks--A New Tool for Discrimination of Eps in a Clinical Study= |
---|---|
作者 | Dickhaus,Hartmut; Heinrich,Hartmut; |
期刊 | 醫學工程 |
出版日期 | 19961200 |
卷期 | 8:6 1996.12[民85.12] |
頁次 | 頁109-123 |
分類號 | 410.1644 |
語文 | eng |
關鍵詞 | Wavelet network; Clinical study; |
英文摘要 | Wavelet networks (WNs) are introduced as a hybrid system approach for discrimination tasks based on nonstationary signal characteristics. Feature extraction being one important step of the procedure is performed by the wavelet transform, an appropriate method for time-frequency representations, whereas classification is carried out by a multi-layer preceptron. The approach is demonstrated as an intelligent self-learning mechanism concerning the network weights as well as the feature parameters. In contrast to other methods suggested in literature feature selection and training of the classifier is performed simultaneously in our approach. This property becomes to an important advantage in real applications. Besides an example with simulated signals the capabilities of the WN method are illustrated by a clinical study. The purpose of this study was to discriminate children with attention deficit disorders with hyperactivity (ADDH) from a control group of healthy children by means of acoustic evoked potentials (AEPs). A selective attention paradigm (SELA) was applied to 25 healthy and 25 hyperactive boys. Because of the high interindividual variability of the recorded EP signals quantification of the averaged sweeps by peak latencies and amplitudes for discrimination failed. However, time-frequency distributions calculated from wavelet transformed responses gave reason to apply our approach of WNs for discrimination between the two groups. Two different network configurations are evaluated. 80% correct classifications was the best value achieved by crossvalidation. The results are considered on the background of the specific clinical situation. The advantages of the WN approach compared with frequently applied methods using time or frequency domain features are discussed. |
本系統之摘要資訊系依該期刊論文摘要之資訊為主。