TABLE III.
Input time series used by different models during the model development for multiclass arrhythmia classification.
| Index | Model name | Model architecture | Model input | Notes | ||||
|---|---|---|---|---|---|---|---|---|
| 1D | 2D | |||||||
| PPG | HR | ACC | Magnified HR |
TFS | ||||
| 1 | 1D-VGG-16 [6](PPG only) | 1D VGG-16 model | ✓ | Retrained Liu et al. model6 on the same data we used. | ||||
| 2 | 1D-VGG-16 [6](four channels) | ✓ | ✓ | ✓ | ✓ | |||
| 3 | 2D DenseNet [8](2D TFS) | 2D DenseNet model | ✓ | Retrained Chen et al.8 model on the same data we used. | ||||
| 4 | PPG only | Our 1D-Bi-GRU model | ✓ | |||||
| 5 | PPG + ACC | ✓ | ✓ | Added accelerometer (ACC) signal. | ||||
| 6 | HR only | ✓ | Heart rate (HR) was calculated from the PPG peaks with WEPD algorithm25 and was normalized with fixed range [30, 220] BPM. | |||||
| 7 | HR + ACC | ✓ | ✓ | |||||
| 8 | HR + magHR | ✓ | ✓ | Magnified HR was normalized with each 30-sec segment’s minimum and maximum HR. | ||||
| 9 | HR + magHR + ACC | ✓ | ✓ | ✓ | ||||
| 10 | PPG + HR | ✓ | ✓ | |||||
| 11 | PPG + HR + ACC | ✓ | ✓ | ✓ | ||||
| 12 | PPG + HR + ACC + magHR (best model) | ✓ | ✓ | ✓ | ✓ | The model with the best performance on our Pulsewatch dataset. | ||