Personal Identification Using Bathtub Electrocardiograms with a Machine Learning Long Short-term Memory Model

Abstract

For health management purposes, a system has been developed to acquire electrocardiograms (ECGs) while bathing in a bathtub equipped with electrodes. The bathers were people who lived in the same house and personal identification is required to classify ECG data of each bather. As with the acquisition of bathtub ECGs, personal identification should be performed unconsciously and unconstrained. Bathing involves thermal and hydrostatic stresses. Bathtub ECG changes with increased heart rate and cardiac output due to bathing. In this study, a machine learning long short-term memory model was used to train an ECG of 100 beats immediately before and after bathing to identify individuals. The results showed a 100% identification rate. To evaluate the robustness of the training model to noise, white noise of 50% and 100% intensity relative to the amplitude of the R- to S-waves was added to the ECG of the training data. Subsequently, the average identification rates reduced to 94.6% and 45.1%, respectively. This suggests that the noise up to about 50% of the intensity relative to the amplitude of the R- to S-wave could maintain a high identification rate.