CNN-LSTM-based models to predict the heart rate using PPG signal from wearables during physical exercise

Abstract

Atrial fibrillation, or AFib is the most common form of arrhythmia, in fact, 3\% of people over the age of 20 suffer from this condition and more shockingly, it is found that patients with arrhythmias are 5 times more likely to have a stroke [1]. These events of irregularity in the heart beat occur briefly and can be very sporadic which leads their detection to be rather cumbersome, with the standard diagnostic procedure being a long term continuous ECG. This leads to multiple problems, first of all, the ECG is commonly performed as the person is laying down in a hospital bed, which immediately distances the test environment from the real world scenario of living with AFib or another kind of arrhythmia, especially since arrhythmias are more likely to manifest during the practice of physical exercise. From this need arises the alternative of using a PPG (Photoplethysmography) signal, which is an optical method of measuring the blood volume in surfaces such as the finger tip, wrist or ear lobe[2] and can be present in many portable devices like fitness bands and smartwatches, therefore enabling it to be used during the practice of physical exercise [3]. This alternative heart rate monitor is substantially less invasive and more mobile but it is also much more susceptible to motion artifacts. However the motion artifacts that create this noise can be quantified through the pairing of an accelerometer to this device, which provides us with data regarding the acceleration of the devices over the 3 axis. Data like this is available and we will be using the dataset from the IEEE Signal Processing Cup 2015, with which, a plethora of different approaches to remove the noise becomes available, from more traditional filtering methods to the more modern Artificial Intelligence approaches, like the neural networks and support vector machines that have been used in the related work. We believe this multimodal approach will provide us with substantially better results than traditional methods that used the signal itself as the only input of the model.