An algorithm for automatic detection of repeater F-waves and MUNE studies
Üst veriTüm öğe kaydını göster
KünyeArtug, N. T., Sirin, N. G., Akarsu, E. O., Baslo, M. B., & Oge, A. E. (2019). An algorithm for automatic detection of repeater F-waves and MUNE studies. Biomedical Signal Processing and Control, 51, 264-276. doi:10.1016/j.bspc.2019.02.025
The present study aims to develop an algorithm and software that automatically detects repeater F-waves which are very difficult to analyze when elicited as high number of recordings in motor unit number estimation studies. The main strategy of the study was to take the repeater F waves discriminated by the neurologist, from limited number of recordings, as the gold standard and to test the conformity of the results of the new automated method. Ten patients with ALS and ten healthy controls were evaluated. 90 F-waves with supramaximal stimuli and 300 F-waves with submaximal stimuli were recorded. Supramaximal recordings were evaluated both manually by an expert neurologist and automatically by the developed software to test the performance of the algorithm. The results both acquired from the neurologist and from the software were found compatible. Therefore, the main expected impact of the present study is to make the analysis of repeater F waves easier primarily in motor unit number estimation studies, since there is currently a continuing need for such automated programs in clinical neurophysiology. Submaximal recordings were examined only by the developed software. The extracted features were: maximum M response amplitude, mean power of M response, mean of sMUP values, MUNE value, number of baskets, persistence of F-waves, persistence of repeater F-waves, mean of F-waves' powers, median of F-waves' powers. Feature selection methods were also applied to determine the most valuable features. Various classifiers such as multi-layer perceptron (MLP), radial basis function network (RBF), support vector machines (SVM) and k nearest neighbors (k-NN) were tested to differentiate two classes. Initially all features, then decreased numbers of features after feature selection process were applied to the aforementioned classifiers. The classification performance usually increased when decreased features were applied to intelligent systems. Ulnar recordings under submaximal stimulation showed better performance when compared with supramaximal equivalents or median nerve equivalents. The highest performance was obtained as 90% with k-NN algorithm which was a committee decision based classifier. This result was achieved with only two features, namely mean of sMUP amplitude and MUNE value.