Table 5.

Comparison of methods and result of state-of-the-art studies with proposed model

	Data source	Method	Personalized	Metrics	SBP	DBP
(Su et al. 2018) [34]	Proprietary data (84 subjects, 10 min each)	PTT + deep learning (LSTM)	Unknown	RMSE	3.73	2.43
(Mohammad Kachuee et al. 2017) [35]	MIMIC II (1000 subjects)	PTT + classical ML (AdaBoost)	Yes	MAE	11.17	5.35
(Kurylyak, Lamonaca, and Grimaldi 2013b) [51]	MIMIC (15,000 beats)	Temporal PPG features + artificial neural network (ANN))	Unknown	MAE	3.8	2.21
(Gupta et al. 2022) [52]	MIMIC I, II, and III (39 subjects, ? subjects, 510 subjects)	PPG signal’s derivative contours + ML algorithms	Yes	MAE	0.74, 1.69, 1.30	0.35, 0.77, 0.56
(Bernard, Msigwa, and Yun 2022) [36]	MIMIC II (69 subjects) and proprietary data (23 subjects)	five 1-D CNN, three Bi-directional LSTM networks	No	MAE	1.38	0.95
(Leitner, Chiang, and Dey 2022) [37]	MIMIC III (100 subjects)	RCNN neural networks + personalization	Yes	MAE	3.52	2.2
(Wang et al. 2022) [53]	MIMIC II (348 records)	Visibility graph + transfer learning	No	MAE	6.17	3.66
(Schlesinger et al. 2020) [38]	MIMIC II (304 subjects)	CNN + Siamese Network	No	MAE	5.95	3.41
Our proposed work (MLM-Transformer w/personalization)	MIMIC III (1,732 subjects)	3-layer Time Series Transformer + personalized fine-tuning	Yes	MAE	2.41	1.31