CNN-RNN architecture to calculate BPM from underwater ECG samples

Abstract

AbstractThis paper presents a novel approach for the generation of heart Beats Per Minute (BPM) from noisy/distorted underwater Electrocardiogram (ECG) samples. We solve this problem using a software based approach via a Convolutional - Recurrent (CNN-RNN) regression model and demonstrate good performance: Mean Absolute Error (MAE): 0.400, Root Mean Square Error (RMSE): 0.653 - for counted underwater heart beats. The neural network is trained on land based ECG samples that have been modified by replicating the signal noise/artefacts seen on under water ECG signals; this process has not yet been reported in literature. This allows the prediction of complex samples without the need for underwater sampling and improves neural network performance. To verify this approach, the trained neural network is tested on underwater ECG samples. This solution requires minimal signal pre-processing and does not require any specific ECG electrode/amplifier design. We have done this to minimise cost and ensure easy deployment. In addition, the techniques discussed here can be applied to any sampled ECG signals and are not hardware specific. This will lead to improved performance where underwater BPM data is required, for example: performance sport; rehabilitation and/or divers in hazardous environments.