Table 3.
Comparison with well-established related work in terms of data used, methodology and errors.
| Author | Data Used | Method Used | Personalization | Error |
|---|---|---|---|---|
| Chan et al. [12] | Unspecified proprietary data | PTT approach, classical ML (linear regression) | Yes | ME of 7.5 for SBP and 4.1 for DBP |
| Su et al. [14] | Proprietary data (84 subjects, 10 min each) | PTT approach, deep learning (long short-term memory (LSTM)) | Unknown | RMSE of 3.73 for SBP and 2.43 for DBP |
| Kachuee et al. [13] | MIMIC II (1000 subjects) | PTT approach, classical ML (AdaBoost) | Optional | MAE of 11.17 for SBP and 5.35 for DBP |
| Teng et al. [17] | Proprietary data (15 subjects, 18 seconds each) | Temporal PPG features, classical ML (linear regression) | Unknown | ME of 0.21 for SBP and 0.02 for DBP |
| Kurylyak et al. [18] | MIMIC (15,000 beats) | Temporal PPG features, deep learning (fully-connected artificial neural network (ANN)) | Unknown | MAE of 3.80 for SBP and 2.21 for DBP |
| Xing et al. [19] | MIMIC II (69 subjects) and proprietary data (23 subjects) | Frequency PPG features, deep learning (fully-connected ANN) | Unknown | RMSE of 0.06 for SBP and 0.01 for DBP |
| Our work | MIMIC III (510 subjects) | Temporal and frequency features of PPG, PPG’ and PPG”, deep learning (spectro-temporal ResNet) | Yes | MAE of 9.43 for SBP and 6.88 for DBP |