Development of a Micro-Macro Neural Network to recognize rollover movement

Bone metastasis patients suffer from pain when their trunks twist during movements such as rollovers. In this overall research project, our ultimate aim is to develop an effective rollover-support system for patients with cancer bone metastasis. The core of this system will be a pneumatic rubber muscle that will be operated based on the EMG signals from the patient's internal abdominal oblique muscle to limit the range of motion of the trunk twist only when the patients will feel the pain. The Time Delay Neural Network (TDNN) is the traditional method for recognizing the movement such as rollover using EMG signals. We have developed a new neural network, called the Micro-Macro Neural Network (MMNN), to recognize the rollover movement earlier and with more accuracy than possible with the TDNN. Recognition using MMNN was 49 (S. D. 45) (msec) faster than that using TDNN. The recognition rate before the rollover started was improved from 38% (TDNN) to 86% (MMNN). Additionally, the number of false recognitions using MMNN fell to only one third of those using TDNN. In addition, by using the unit contribution rate of the neural network, we found that the MMNN effectively accounted for the importance of past EMG data (the data gathered before the current measurement point). We also found that the de-noising performance of the MMNN was effective.