Estimating Heart Rate from Seismocardiogram Signal using a Novel Deep Dominant Frequency Regressor and Domain Adversarial Training

Abstract

Estimating heart rate (HR) from the seismocardiogram (SCG) signal can provide a more unobtrusive alternative for long-term HR monitoring where the gold standard electrocardiogram (ECG) signal is less favorable. Deep learning (DL) approaches have demonstrated promise in accomplishing this and are attractive due to their flexible and data-driven nature. However, current dense layer-based DL approaches lack a carefully designed regressor to estimate the HR in the SCG input and may overfit in low-data regime. It is also uncertain how well most of these DL approaches can generalize to unseen subjects, as evaluation has primarily taken place with no separation of subjects between training and testing data. In this work, we address these limitations by designing a DL approach for SCG HR estimation that leverages our proposed Dominant Frequency Regressor (DFR) with a Fast Fourier Transform layer and test the model performance using leave-one-subject-out cross validation. Specifically, we measure ECG and SCG from 19 subjects using our custom-built wearable patch. Here, ECG is used for training and regularization only, and SCG is used for training and inference. These signals are bandpass filtered, segmented into 60s windows, and normalized. Next, the proposed DFR-based DL model is applied and regularized by domain adversarial training. We report a mean absolute error (MAE) of 1.42±1.66 beats per minute (bpm) for HR with a range of 63-104 bpm. The proposed method can augment existing methods or be adapted for other problems of similar nature owning to its superiority to the dense layers-based alternatives. This work can lead to accurate, real-time algorithms for estimating 60s HR from a single-chest worn accelerometer only, which could be embedded in textiles.