An interpretable multi-scale lightweight network for patient-ventilator asynchrony detection during mechanical ventilation

•Combining convolution with self-attention mechanism into a fusion network for PVA detection.•Local features and long-distance relevant multi-scale information of waveform images are automatically extracted simultaneously.•A novel lightweight parallel structure and quantization method are proposed to reduce the parameters and computational consumption of the deep neural network.•Relative position coding and interpretability module are established to visualize the classification while enhancing the spatial association and improve the understanding of the model's decision-making.•The average identification accuracy, sensitivity and F1 scores of the model are 0.987, 0.983, and 0.989, respectively. Patient-ventilator asynchrony (PVA) is commonly the result of an inconsistency between the patient's respiratory rhythm and the ventilator rate during mechanical ventilation, which will lead to serious prognostic problems. Therefore, it is necessary to develop a reliable automatic recognition algorithm. However, many current intelligent algorithms, suffer from poor detection efficiency due to the limited datasets, lack of interpretability, with huge amounts of parameters, which are difficult to deploy on resource-constrained hardware devices. In this paper, an interpretable multi-scale lightweight network, Mobiformer, is proposed for automatic PVA identification. The model adopts a novel lightweight parallel structure, combining the ability of parameter sharing of convolutional network and the global interaction of self-attention mechanism for multi-level feature extraction. It can dynamically capture the global dependencies while extracting the local information of the waveforms, thus detecting the occurrence of PVA more accurately. In addition, a relative position encoding method is employed and a visualization module is built to improve the classification accuracy while making the result interpretable, avoiding the black-box problem. In the test, the identification accuracy, sensitivity, specificity, and F1 scores of the model are 0.987, 0.983, 0.986, and 0.989, respectively. Moreover, the critical waveform components of different categories are highlighted, consistent with the understanding of PVA by physicians. The experimental results show that our model can identify PVA more effectively and make the decision interpretable, which offers tremendous potential for future clinical applications.