Fig. 5: Long-term monitoring of respiratory status and comparison of the performances of the two models.

a–d Continuous collection of 32-s pulse data from volunteers using a flexible microsystem, with their breathing states calculated and compared against those monitored using commercial breathing sensors and the two models (ResNet-BiLSTM and BiLSTM). e A radar chart illustrating the mean classification accuracy of different models when analyzing the respiratory patterns of six volunteers. The larger the area of the graph, the higher the model’s average accuracy; the greater the symmetry, the more robust the model. f Dimensionality reduction of the output features from the “bilstm1” layer of the ResNet-BiLSTM model using the t-SNE technique. g Dimensionality reduction of the output features from the “bilstm1” layer of the BiLSTM model using the t-SNE technique