This study summarizes the application of automatic recognition technologies for patient-ventilator asynchrony (PVA) during mechanical ventilation. In the early stages, the method of setting rules and thresholds relied on manual interpretation of ventilator parameters and waveforms. While these methods were intuitive and easy to operate, they were relatively sensitive in threshold setting and rule selection and could not adapt well to minor changes in patient status. Subsequently, machine learning and deep learning technologies began to emerge and develop. These technologies automatically extract and learn data characteristics through algorithms, making PVA detection more robust and universal. Among them, logistic regression, support vector machines, random forest, hidden Markov models, convolutional autoencoders, long short-term memory networks, one-dimensional convolutional neural networks, etc., have all been successfully used for PVA recognition. Despite the significant advancements in feature extraction through deep learning methods, their demand for labelled data is high, potentially consuming significant medical resources. Therefore, the combination of reinforcement learning and self-supervised learning may be a viable solution. In addition, most algorithm validations are based on a single dataset, so the need for cross-dataset validation in the future will be an important and challenging direction for development.
该研究总结了患者与呼吸机非同步性（patient-ventilator asynchrony，PVA）自动识别技术在机械通气过程中的应用。在早期阶段，规则及阈值的设定方法依赖于呼吸机参数及波形的人为解析，虽然这类方法直观并易于操作，但在阈值设定和规则选择上相对敏感，不能很好地适应患者状态的微小变动。随后，机器学习和深度学习的技术开始出现并发展。这些技术通过算法自动提炼和学习数据特性，使PVA的检测更具鲁棒性和通用性。其中，逻辑回归、支持向量机、随机森林、隐马尔可夫模型、卷积自编码器、长短期记忆网络、一维卷积神经网络等方法都被成功地用于PVA的识别。尽管深度学习方法在特性提取上取得了显著进步，但是它们对标签数据的需求较大，可能会消耗大量医疗资源。因此，强化学习与自监督学习的结合可能是一个实际可行的解决方案。此外，算法的验证大多基于单一的数据集，未来对于跨数据集验证的需求将是一个重要且充满挑战性的发展方向。.
Keywords: algorithm; automatic detection; mechanical ventilation; patient-ventilator asynchrony.