Table 11.
Comparison of past studies based on MAE performance.
| Study | Year Published | Dataset | Input Signals | Method | MAE (mmHg) | |
|---|---|---|---|---|---|---|
| SBP | DBP | |||||
| Kurylayak et al. [18] | 2013 | MIMIC, 15,000 Pulsations | PPG | ANN | 3.80 | 2.21 |
| Wang et al. [19] | 2018 | MIMIC, 72 Subjects | PPG | ANN | 4.02 | 2.27 |
| Slapničar et al. [24] | 2019 | MIMIC, 510 Subjects | PPG | CNN | 9.43 | 6.88 |
| Miao et al. [40] | 2019 | 1711 ICU and 30 Arrythmia Patients | ECG | CNN + LSTM | 7.10 | 4.61 |
| Esmaelpoor et al. [39] | 2020 | MIMIC-II (200 Subjects) | PPG | CNN + LSTM | 1.91 | 0.67 |
| Ibtehaz et al. [13] | 2020 | MIMIC-II (942 Subjects) | PPG | CNN + CNN | 5.73 | 3.45 |
| Li et al. [22] | 2020 | MIMIC-II (3000 Records from UCI Repository) | PPG, ECG | LSTM | 4.63 | 3.15 |
| Hsu et al. [21] | 2020 | MIMIC-II (9000 Records from UCI Repository) | PPG, ECG | ANN | 3.21 | 2.23 |
| Athaya et al. [25] | 2021 | MIMIC-II (100 Subjects) | PPG | CNN | 3.68 | 1.97 |
| Harfiya et al. [23] | 2021 | MIMIC-II (5289 Records from UCI Repository) | PPG | LSTM | 4.05 | 2.41 |
| Baker et al. [65] | 2021 | MIMIC-III | PPG, ECG | CNN + LSTM | 4.41 | 2.91 |
| Rong et al. [66] | 2021 | MIMIC-II (UCI Repository) | PPG | CNN + LSTM | 5.59 | 3.36 |
| Sagirova et al. [64] | 2021 | 512 Patients | ECG, PPG | |||
| Qin et al. [41] | 2021 | MIMIC-II (1227 Records from UCI Repository) | PPG | VAE | 7.95 | 4.11 |
| This Study | 2021 | MIMIC-II (942 Subjects–12,000 Recordings from UCI Repository) | PPG, ECG | CNN + ANN | 2.333 | 0.713 |
| MIMIC-II + BCG (942 + 40 = 982 Subjects) | 2.728 | 1.166 | ||||
| AAMI Standard | ≤ | |||||
Note: It is important to mention that Hsu et al. [23] in their paper reported that they used 9000 subjects’ data for BP prediction from the UCI repository, but it was 9000 out of 12,000 instances or recordings of data collected from the MIMIC-II dataset. These are the data from 942 patients as reported by Kachuee et al. [14], the originator of this dataset. A similar occurrence happened for the case of Harfiya et al. [25] where they reported 5289 signal instances from the UCI repository as 5289 patients. In comparison, this study fully utilized all 12,000 instances.