Respiratory pattern prediction for synchronized functional electrical stimulation

Abstract

In addition to physical, psychological and social problems, paralysis in the ventilatory muscles can be one of the side effects caused by spinal cord injury. Due to the total or partial impediment of the conduction of the action potentials that perform the muscle activation, there may be changes in the ventilatory mechanics, affecting the body’s ability to maintain oxygenation due to diaphragmatic paralysis. In addition, coughing events, which are important for secretion removal, may be hampered by paralysis of the abdominal muscles. This problem is the main cause of death in people with spinal cord injury. Transcutaneous functional electrical stimulation, automatically synchronized with spontaneous breathing is considered an alternative technique for respiratory rehabilitation in people with spinal cord injury. However, during stimulation, other events may occur, such as coughing, body movement, speech, among others, which need to be treated individually by the stimulation system. The main goal of this research is to develop a machine learning model for automatic prediction of respiratory patterns for synchronized electrical stimulation systems. Different types of sensors were evaluated to acquire the respiratory signal: differential pressure sensor connected to a pneumotachometer; strain gauges in a load cell structure coupled to an elastic band, to measure chest expansion during breathing; and EMG electrodes positioned in the regions of the sternocleidomastoid, external oblique and omohyoid muscles, to detect cough. Data from 20 healthy volunteers were acquired during voluntary events of coughing, normal breathing, speech and body movement, and a tagged dataset was created. A machine learning algorithm based on Long Short-Term Memory Networks was developed to classify the data between the classes “precough”, “normal breathing” and “interfering events (speech and body movement)”. The accuracy was 80.24%, using the combination of data from the pressure sensor and the electrodes positioned on the omohyoid muscle. The developed algorithm is generalizable, allowing use in people who did not participate in the training phase of the network. In this way, it becomes a strong candidate to be used in synchronized electrical stimulation systems. However, the pressure sensor connected to the pneumotachometer is considered an intrusive sensor, and not comfortable for prolonged use.