External respiratory signals prediction models

The LSTM prediction (LSTMpred) and SVR prediction (SVRpred) models were established using the LSTM neural and SVR networks, respectively, to predict the respiratory motion of the chest surface at the system latency. Both LSTMpred and SVRpred use the current external respiratory motion signals S_t to predict the future external respiratory motion signals S_t+i. Considering that the length of the training and label datasets should be the same during the training of the SVR model, a S_t-to-S_t+i length ratio of 1:1 was selected for both models. As the typical system latency ranges from tens of milliseconds to more than 400 ms, the respiratory motion prediction algorithm results were assessed at latencies of 50, 150, 200, and 450 ms, i.e., the i = 1, 3, 4, and 9 latencies were set to 50, 150, 200, and 450 ms, respectively.

The external respiratory motion data of dataset D_v were divided at ratios of 9:1 into training and testing data for the LSTMpred and SVRpred models. After testing, comparison, and adjustment to determine the best network performance, appropriate parameters were selected for the two external prediction models. For the LSTMpred model, mean square error (MSE) loss was chosen as the loss function and the network was trained using an Adam optimizer with a learning rate of 0.001 and a batch size of 64. A dropout rate of 20% over 60 total epochs and 20 time steps was used in the LSTMpred model adjustment process. The number of layers and neurons in each layer are shown in Fig. 1. For the SVRpred model, the radial basis function (RBF) kernel was selected as the kernel function, and the gamma and penalty parameter C of SVRpred were set as 0.1 and 1000, respectively.

Structure of LSTMpred and LSTMcorr models