Larynx Function prediction using tunable-Q wavelet transform based filter bank using Surface EMG Data

People with total laryngectomy (TL) find difficulties in speaking, swallowing, speaking, and coughing. Their quality of life is hugely impacted as the larynx is unable to perform critical functions. In this study, we analyze surface electromyography (EMG) signals from the larynx of both control participants and TL patients, encompassing four actions namely swallow, cough, speech, and random movements. Data processing focused on the EMG-intercostal, EMG-submental, and EMG-diaphragm signals, adopting a frame size of 128 ms with a 64 ms overlap. A tunable-Q wavelet transform (TQWT) based filter bank was applied to decompose the signals into 40 elements, and signals with more than 20% of the total signal energy were re-merged into a filtered signal. The proposed method is evaluated on an open-source data set. Its efficacy in classifying swallow, cough, speech, and null is evaluated for control and TL subjects. Feature extraction yielded two frequency-based features, namely mean frequency and median frequency, and six time-domain features, namely Log Teager Kaiser Energy Operator, Mean Absolute Value, Slope Sign Change, Zero Crossing, Waveform Length, and Willison Amplitude. Our analysis utilized a Weighted K-nearest neighbors (WKNN) approach to classify the recordings, achieving F1-scores of 82.82% (swallows), 68.15% (coughs), and 76.00% (speech) for controls and 53.30% (swallows), 63.95% (coughs) and 76.19% (speech) for Total Larengotomy (TL) group, using WKNN classifier. With a significant improvement in the control group’s swallow and speech F1 scores, a marginal improvement in speech F1 score for the TL group. Using WKNN, the classification accuracy achieved in classifying swallow, cough, and speech for control and TL group is 79.8% and 77.03%, respectively.