Self-adapting Classification System for Swallow Intention Detection in Dysphagia Therapy

Abstract

In dysphagia the ability of elevating the larynx and hyoid is usually impaired. Electromyography (EMG) and Bioimpedance (BI) measurements at the neck can be used to trigger functional electrical stimulation (FES) of swallowing related muscles. The height and speed of larynx elevation can be assessed by evaluating the BI during a swallow. For the triggering of an supporting FES and for biofeedback online detection of swallow onsets is required. Patients can practice by a gamified biofeedback to swallow harder, swallow in a timely manner or to maintain the larynx elevation for a longer time period (Mendelson maneuver). The success of the stimulation and biofeedback therapy as well as the acceptance by the patient strongly depends on the precise detection of swallow onsets. We have introduced a classification algorithm based on a random forest classifier to trigger FES in phase with voluntary swallowing based on EMG and BI. Although the classification is successful in healthy subjects, difficulties appear in the utilization on some patients. The reason for this can be found in a strongly varying residual swallow activity. Usually the activity of EMG and change in BI are smaller in patients compared to healthy subjects. Thus an adaption procedure is needed, that can be easily applied. In this paper we introduce an algorithm that is capable to find an optimal classifier for a patient in terms of sensitivity. The adaption algorithm uses a small number of recorded swallow onsets of a patient at the beginning of a therapy session to evaluate different classifiers and to pick the most suitable for the treatment. The set of random forest classifiers has been trained with data from healthy subjects by step wise shifting the class weights of swallows and non-swallows, yielding classifiers with different sensitivities. The evaluation is done using data from 41 patients. It showed that the sensitivity of the classification can be increased by 4 to 6 % in average compared to fixed classifiers, but up to 66 % for individual patients. Finally, we studied the effect this adaptive classifier in triggered stimulation therapy in a single dysphagia patient. Swallowing performance was measurements during one week of therapy consisting of eleven therapy session. An improvement of 40 and 63 % in larynx elevation and speed could be observed, respectively.

Keywords

• dysphagia
• biofeedback
• classification