RRWaveNet: A Compact End-to-End Multi-Scale Residual CNN for Robust PPG Respiratory Rate Estimation

TL;DR

This study proposes a method for continuously estimating RR, RRWaveNet, a compact end-to-end deep learning model which does not require feature engineering and can use low-cost raw photoplethysmography as input signal and shows feasibility of remote RR monitoring in the context of telemedicine and at home.

Abstract

Respiratory rate (RR) is an important biomarker as RR changes can reflect severe medical events such as heart disease, lung disease, and sleep disorders. Unfortunately, standard manual RR counting is prone to human error and cannot be performed continuously. This study proposes a method for continuously estimating RR, RRWaveNet. The method is a compact end-to-end deep learning model which does not require feature engineering and can use low-cost raw photoplethysmography (PPG) as input signal. RRWaveNet was tested subject-independently and compared to baseline in four datasets (BIDMC, CapnoBase, WESAD, and SensAI) and using three window sizes (16, 32, and 64 seconds). RRWaveNet outperformed current state-of-the-art methods with mean absolute errors at optimal window size of 1.66 \pm 1.01, 1.59 \pm 1.08, 1.92 \pm 0.96 and 1.23 \pm 0.61 breaths per minute for each dataset. In remote monitoring settings, such as in the WESAD and SensAI datasets, we apply transfer learning to improve the performance using two other ICU datasets as pretraining datasets, reducing the MAE by up to 21$\%$. This shows that this model allows accurate and practical estimation of RR on affordable and wearable devices. Our study also shows feasibility of remote RR monitoring in the context of telemedicine and at home.

Figures (8)

• Figure 1: Deep neural network architectures for respiratory rate estimation from recent studies: (a) Bian et al. bian applies ResNet blocks and tunes their hyperparameters using Bayesian optimization, (b) respwatch, from Dai et al., adapts the convolutional neural network to help capture the motion artifcats in the PPG signal, and (c) Ravichandran et al. respnet proposes an IncResU-Net-like architecture consisting of multiple encoders and decoders.
• Figure 2: Our proposed architecture, RRWaveNet. Composed of three modules, RRWaveNet involves the multi-scale convolution (left), the deep spatial-temporal residual blocks (center), and the respiratory rate estimator (right). Each layer's title is abbreviated at the top row for simplicity and the shape of the output tensor after each layer is specified below its title. For example, conv1dk32s5, the leftmost layer in the center module, refers to a 1D-convolutional layer with a kernel of size 32 and a stride of 5, resulting a $(10W, 1)$ tensor.
• Figure 3: For performance evaluation, our study uses data from various data sources with different characteristics. WESAD and SensAI are noisy but large datasets obtained using a wrist-worn PPG sensor. In contrast, the BIDMC and CapnoBase datasets are higher quality datasets that used pulse oximeters attached to the fingertip.
• Figure 4: Our evaluation method as used in both experiments is based on the leave-one-out validation method which ensures subject independence. All patient samples, labelled with distinct colors, are one-dimensional with length of $S_RW$. As sequentially labelled using numbers above, first, samples from one patient are left as the test set, while those from the remaining patients belong to the training and validation set. The latter is further divided with a ratio of 4:1 to train the model with a five-fold cross validation. Each fold evaluates the test samples and outputs one MAE, which is averaged over the five folds. Each patient takes turn as the test set.
• Figure 5: The distribution of respiratory rate containing 4000 samples from each dataset. The CapnoBase dataset includes respiratory rates that are higher than 35 BPM, which do not appear in other three datasets.