Decentralised biomedical signal classification using early exits

This paper presents a decentralised signal classification approach for data acquired using Internet of Things (IoT) wearable sensors. Traditionally, data from IoT sensors are processed in a centralised fashion, and in a single node. This approach has several limitations, such as high energy consumption on the edge sensor, longer response times, etc. We present a distributed processing approach for convolutional neural network (CNN) based classifiers where a single CNN model can be split into multiple sub-networks using early exits. To reduce the transfer of large feature maps between subnetworks, we introduced an encoder-decoder pair at the exit points. Processing of inputs that can be classified with high confidence at an exit point will be terminated early, without needing to traverse the entire network. The initial sub-networks can be deployed on the edge to reduce sensor energy consumption and overall complexity. We also experimented with multiple exit point locations and show that the point of exit can be adjusted for trade-offs between complexity and performance. The proposed system can achieve a sensitivity of 98.45% and an accuracy of 97.55% for electrocardiogram (ECG) classification and save 60% of the data transmitted wirelessly while reducing 38.45% of the complexity.