Table 4.

Overall prediction performance (measured as MAE) of the proposed Transformer Models (bottom 3) on the test dataset for ASBP, ADBP, and SpO₂ in comparison with traditional Transformer and U-Net model as baseline. MLM, fine-tuned with Masked Language Modeling

Methods	Parameter Counts, × 106	SpO2, %	ASBP, mmHg	ADBP, mmHg
Transformer	1.4	1.65 ± 1.94	6.76 ± 5.24	3.57 ± 4.39
U-Net	13.3	1.02 ± 1.46	5.03 ± 4.78	2.98 ± 3.41
Transformer w/multi-task	2.2	1.28 ± 1.67	6.44 ± 5.32	3.42 ± 4.17
MLM-Transformer	2.2	0.75 ± 1.04	4.97 ± 4.72	2.99 ± 2.39
MLM-Transformer w/personalization	2.2	0.56 ± 0.79	2.41 ± 2.72	1.31 ± 1.77