Integrating Differential Signals Into RNNs for Real-Time Respiratory Prediction

In lung nodule puncture and abdominal radiotherapy, the motion of lesions caused by patient respiration can compromise procedural accuracy. Consequently, integrating respiratory prediction can significantly enhance surgical precision. However, traditional methods often fail to provide real-time capabilities and may inadequately represent the dynamic characteristics of respiratory signals, resulting in potential inaccuracies and inefficiencies. In this study, we aimed to develop a novel real-time respiratory prediction model that incorporates respiratory signals along with their first- and second-order derivatives within a recurrent neural network framework. By incorporating differential signals, our model effectively learns the high-frequency characteristics of respiratory signals, enhancing its sensitivity to short-term respiratory fluctuations while preserving the ability to capture long-term dependencies. This approach addresses the limitations of current methods in handling nonlinear, quasi-periodic, and nonstationary respiratory signals. Our model achieved superior performance across prediction windows of 200, 400, and 600 ms, with mean absolute errors of 0.026, 0.216, and 0.394 and root mean squared errors of 0.034, 0.288, and 0.534, respectively. This study enhances the precision of respiratory prediction and delineates the critical role of differential signals in forecasting quasi-periodic physiological patterns.