Table 2 An overview of HR tracking techniques
From: Heart rate tracking in photoplethysmography signals affected by motion artifacts: a review
Tracker | Classifier | Tracking method | HR Stages | Specialty | Error1 | Error2 |
|---|---|---|---|---|---|---|
– | SDT | Initialization, peak selection, peak verification, and Smoother algo. (JOSS) | Search areas for F0 and F1 in PPG spectra | 2.42 and 1.28 | 1.82 and 1.01 | |
CC [51], 2015 | – | SDT | Heuristic | Two spectral peaks in PPG and considers direction and direction of error | 1.83 | 1.4 |
Emroz et al. [52], 2016 | – | SDT | Comparison of heart rate from multiple (4) sources | Building of HR space, comparison of two sources for HRC at a time, and fine tuning | 1.02 | – |
CNN | SST | Heuristic and trained CNN using 4 channel images converted from signals | Downsampling of PPG and ACC. Three spectral peaks in PPG and ACC and iterative selection of spectral peak and end-to-end learning | 1.93 and 3.56 | 2.07 and – | |
CARMA [85], 2015 | – | SST | Heart rate modeling using prior knowledge | Downsampling of PPG and ACC. Three spectral peaks in PPG and ACC and iterative selection of spectral peak | 2.26 | 3.63 |
HSUM [69], 2018 | – | SST | PPG and ACC modeling using HSUM | Automatic detection of heart rate after motion modeling | 0.74 | 0.83 |
Schack et al. [70], 2017 | – | SPT | Maximization of cost function | Downsampling of PPG and ACC signals, cross correlation, and Gaussian filtering, cost function-based tracking | 1.32 | – |
Torres et al. [87], 2016 | – | SPT | HR selection, verification, and BW adjustment | Computationally less expensive hence suited for real-time application | 1.36 | 1.05 |
Timko et al. [76], 2015 | – | SPT | Phase Vocoder & HR history tracking and smoothing | A few parameters to tune hence is useful in real time | 1.05 | 2.23 |
Galli et al. [61], 2018 | – | SPT | Prediction, estimation, and update | Use of KF to calculate heart rates from DFT | 2.45 | 2.21 |
Rocha et al. [96], 2020 | DNN | MLT | Two layers of binary CNN and binary LSTM | FPGA implementation of binary Cornet | 3.75 | – |
Zhu et al. [71], 2019 | NN | MLT | NN modeling, smoothing, and linear regression | Polynomial approximation of false heart rate | 1.03 | 0.79 |
Roy et al. [101], 2018 | MLP-ANN | MLT | Trained NN using features generated by auto encoder from PCA | Generate HR template using clean PPG | 1.47 | 1.1 |
Biswas et al. [20], 2019 | DNN | MLT | CNN extracted features used to LSTM | Two layers of CNN and LSTM | 1.96 | 1.47 |
Sun et al. [74], 2016 | SVM | MLT | SVM trained on peak amplitude and peak-to-peak separation | – | 1.78 | 1.57 |