Methodology of sequential classification of non-invasive multichannel biosignals, oriented to automatic diagnosis of dysphagia

Abstract

Swallowing is a complex process that involves sequential voluntary and involuntary muscle contractions. Malfunction of swallowing related muscles could lead to dysphagia. The Videofluoroscopic Swallowing Study is the current gold standard in dysphagia assessment, but is related to high cost, long wait times, and harmful radiation risk. There is a lack of standardized and non-invasive methods that help to improve the diagnosis and ambulatory care. The visual inspection is a widely used method for evaluating the surface electromyography signal (sEMG) during swallowing, a process highly dependent of the examiners expertise. It is desirable to have a less subjective technique for the automatic detection of normal and abnormal neuromuscular patterns produced during the swallowing process. The current master’s thesis proposes a methodology of classification that allows to detect normal and abnormal muscular sequences related to swallowing using machine learning algorithms. Thus, 22 healthy subjects and 22 patients with dysphagia were recruited to assess neuromuscular activity during the execution of swallowing tasks. A total of 15 features in time, frequency and time-frequency domains were extracted from seven sEMG channel using the sliding window method. Four statistical moments were computed over the estimated EMG features of each channel in order to characterize the neuromuscular sequences executed by the control and dysphagic groups. The optimal combination of sEMG features was computed according to the F1 score. Furthermore, different combinations of channels were assessed. A support vector machine was used as classifier. Its hyperparameters were optimized with the subset of features, and it achieved a F1 score close to 90%. The proposed scheme is the first machine learning approximation for the automatic detection of dysphagia using multichannel sEMG signals.