One particular field of interest within this medical context is the automatic identification of voice disorders, resulting in perceivable hoarseness. Commonly, for this purpose audio recordings of the acoustical voice signal are analyzed with specialized software quantifying the amount of perturbation (noise) in the signal. However, this type of acoustical analysis does not allow for the clear assignment of certain clinical pictures to a distinct set of perturbation parameters. A more revealing approach for voice diagnosis consists in endoscopic examination of the sound-producing vocal folds in the larynx by means of digital high-speed cameras. These cameras are capable of recording the laryngeal movements at a frame rate of several thousand images per second, and thus, allowing for conclusive real-time analysis. However, the task of manually analyzing the resulting high-speed videos is time-consuming and error-prone. Through automated feature extraction from the recordings and subsequent machine learning analysis, laryngeal movement patterns can be quantitatively captured and automatically classified according to different diagnostic classes (e.g. organic and functional dysphonia). By means of the distributedDataMining infrastructure, we evaluated a large number of machine learning paradigms (e.g. Support Vector Machines, Artificial Neural Networks) and corresponding parameter optimization strategies (e.g. Grid search, Evolution strategy, Genetic algorithms). This preliminary evaluation step allowed us to identify certain learning schemes and parameters which are particularly suited for the considered clinical classification task. Details on the proposed methodology and the obtained classification results can be found in [1],[2] and [3].

References