Seismocardiogram Based Respiratory State Detection using Edge-ML on a Microcontroller

In this paper an Edge-Machine learning based method is proposed for detecting respiratory state. The proposed method detects whether a subject is breathing normally or has stopped breathing. A computationally efficient single feature based machine learning model is trained on a PC and the trained model is then deployed on a microcontroller board to perform the respiratory state detection. The modality used for this purpose is called SCG(Seismocardiogram). It is the recording of 3-dimensional motion of the chest wall in terms of acceleration. The SCG signal contains both cardiac and respiratory information in it. And it is modulated by respiratory effort. The respiratory information in the SCG signal is utilized to detect the respiratory state. The machine learning classifier deployed on the microcontroller predicts the respiratory state of a subject by processing the incoming SCG signal. Only the Z-axis signal of the SCG is used for the prediction. The microcontrollers usually have very modest processing power and limited memory. Therefore it is very important to keep the model size low while also requiring less computing power to run it. The proposed single feature based SVM machine learning classifier, which after training is only 14 kb in size, achieves an accuracy of 92 percent. It is then deployed into an ESP-32 microcontroller for inference. The deployed classifier on the microcontroller is able to predict the respiratory state in under a millisecond.