基于相空间重建-卷积神经网络识别混合机械通气人机不同步

机械通气过程中人机不同步(patient-ventilator asynchrony，PVA)是常见问题。随着智能生理闭环通气成为呼吸机的主要发展方向，机械通气过程将不再局限于传统的通气模式，且针对 PVA 的智能识别模型，现有技术存在弱泛化性和高复杂度的特点。为此，该文的主要工作是：首先，将定压型与定容型通气模式的混合作为样本；其次，设置 Hold-out 和留一法两种交叉验证实验，以验证混合通气模式下 PVA 的识别任务可行性。此外，为提高 PVA 识别任务中模型的泛化性能，该文提出了基于相空间重建的卷积神经网络(phase-space reconstruction-based convolutional neural network，PSR-CNN)模型，通过交叉验证对现有公开报道的方法做模型选择。在模型构造过程中，相空间重建的最优时间延迟参数和嵌入维度分别使用平均互信息和伪近邻算法估计；在交叉验证过程中，同时使用降采样和补零技术，以保证实验的正常运行。结果显示，就全局指标 accuracy 和 F1-score 而言，该文提出的 PSR-CNN 模型，分别高出对比模型约 7% 和 6%；且 PSR-CNN 单个样本的平均训练耗时最短，约 2 ms。综上所述，该文探索了混合通气模式下 PVA 识别的可行性，且在该任务的框架内提出了 PSRCNN 模型，提高了 PVA 分类任务中模型的泛化性能，降低了模型的复杂度。该文的工作对呼吸机在工程上的智能化发展具有重要参考意义与应用价值。

Identification for Patient-Ventilator Asynchrony under Hybrid Mechanical Ventilation Based on Convolutional Neural Network with Phase-Space Reconstruction

Patient-ventilator asynchrony (PVA) commonly occurs during mechanical ventilation. Considering the developing trend of physiological loop ventilation and weak generalization and high complexity of public methods, this paper firstly mixes different ventilation modes simultaneously as sample, and then two cross validations, Hold-out and Leave One Subject Out, are introduced to verify the feasibility of the task that classifying PVAs under hybrid ventilation modes. To solve the drawback of current models, the phase-space reconstruction-based convolutional neural network (PSR-CNN) model is proposed. During model selection, zeropadded and down sampling are applied in order to ensure that all experiments could be conducted smoothly. Results suggest that the performances of PSR-CNN have a higher accuracy and a F1-score than other algorithms. In addition, PSR-CNN shows a shorter time with regard to average training consumption. Overall, this study indicates that the proposed model has a stronger generalization and a decrease in the complexity, which shows application value and provides reference for the intelligent promotion of ventilators in engineering.