Man-machine asynchronous recognition method based on small data set and convolutional neural network

Abstract

The invention relates to a mechanical ventilation man-machine asynchronous recognition method based on a small data set and a convolutional neural network, and the method comprises the steps: converting an original breathing signal into a two-dimensional image, and training a high-precision classification model based on a multi-classification public data set of the two-dimensional image; then freezing weight parameters of layers above a full connection layer in the trained model, and inputting a local respiration waveform image into the model in a transfer learning mode to finely adjust the weight parameter of the last full connection layer; and making the actually-collected original respiration signals preprocessed and then input into the fine-adjusted deep learning model to obtain a man-machine asynchronous recognition result of the current respiration signals. According to the method, a model architecture with the two-dimensional convolutional neural network as a core is adopted, and in a transfer learning mode, under the condition of a small data set, high-precision automatic detection and recognition of man-machine asynchronous waveforms can be achieved, and good interpretability is achieved.

Description

technical field [0001] The invention relates to a two-dimensional convolutional neural network-based recognition method for man-machine asynchrony of mechanical ventilation under a small data set. Belongs to the technical field. Background technique [0002] During mechanical ventilation, man-machine dyssynchrony often occurs. Research data show that the proportion of patients with serious human-machine asynchronous problems is estimated to be between 12% and 43%. However, in clinical practice, doctors can only identify the type of human-machine dyssynchrony based on the pressure and flow time waveforms on the ventilator, and then rely on their own experience. Nowadays, the methods for human-computer asynchronous recognition mainly include three categories of methods, namely, human visual detection, rule-based algorithms, and machine learning-based methods. These three types of methods all have a common feature, that is, the distinguishable feature information in the resp...

AI Technical Summary

Problems solved by technology

This type of algorithm only needs to do simple preprocessing on the waveform, without the need for cumbersome feature engineering construction and selection, and then trains a deep neural network, and then performs recognition. The detection accuracy is high and the robustness is strong, but the limitation is that it requires a large number of training data

However, data labeling is a tedious and time-consuming task

Images