Combining Artificial Intelligence and Automatic Analysis to Study of the Closure Characteristics of Healthy Human Vocal Cords during Phonation Using Synchronous Electroglottography and Laryngeal High-Speed Videoendoscopy

Objective: To use an artificial intelligence neural network to automatically analyze images from high-speed videoendoscopy and synchronous electroglottographic signals in healthy vocal folds to obtain the accurate glottal opening and closing instants in the glottal cycle. This method will be compared to the traditional electroglottography point derivation method to explore the glottal opening and closing characteristics of the vocal folds and better interpret the opening and closing instances of the electroglottography. Methods: Images from high speed videoendoscopy (HSV) and signals from an electroglottogram (EGG) were simultaneously collected for 20 subjects when pronouncing the vowel /i/. High-speed videoendoscopy was used to create a periodic one-dimensional waveform diagram of the glottal area during phonation which was used to calibrate the original EGG signal and obtain the glottal closing instant and the glottal opening instant. This data was combined into a time domain change diagram and used as a training model of the neural network. A prediction of the closed point coordinate values for a segment of EGG signals were calculated and compared against the real value given by videoendoscopy. This process trains the network model to predict a wide range of EGG signals. Results: 1) The positions of the opening and closing instants of the vocal cords identified by the DEGG method were closer to the peaks than those identified by the analysis of the HSV using Glottal Image Explorer software on the same periodic waveform, and the positions predicted by the neural network model were very close to those identified by the calibration method; 2) The data analyzed using the derivative EGG method had an average OQ was 52.27% and an average CQ was 49.52%. In comparison, the data analyzed using the neural network had an average OQ of 34.88% and an average CQ is 66.12%. As such, the EGG analyzed using the neural network was significantly closer to the calibration method, which had an average OQ was 34.60% and an average CQ was 65.42%. Conclusion: There is a high consistency between the glottal opening and closing instants predicted by the trained neural network and the HSV, which can identify and predict the position of the glottal opening and closing points on EGG. This method can provide a more convenient and accurate method for EGG signal analysis and can be applied to the automatic analysis of vibration patterns of various voice diseases.