Adaptive spectral-inertial layer based on a trainable Fourier transform mask for output data filtering in physiological time series forecasting problems

A method for adaptive filtering of output signals of neural network models designed to predict physiological time series is proposed. The developed approach combines spectral filtering based on the fast Fourier transform with a trainable amplitude mask and temporal smoothing implemented through a firstorder exponential filter applied to the output forecast of the recurrent neural network. This combination allows for effective elimination of high-frequency artifacts that arise during abrupt changes in conditions, while preserving the key characteristics of the signal that reflect real physiological dynamics. The practical significance of the developed layer was assessed by a series of experiments on the problem of predicting pulmonary ventilation of rescuers and industrial personnel when passing a route with a given physical load using an isolating breathing apparatus. A recurrent model based on long short-term memory, a type of recurrent neural network architecture trained using individual physiological and route data, was used as a predictive model.