DeltaMamba: Pretrained Differential Mamba for Cuffless Blood Pressure Estimation From Photoplethysmography (PPG)

Photoplethysmography (PPG) provides a practical route for cuffless blood pressure (BP) estimation. Reliable estimation is still difficult because PPG signals vary across subjects, recording conditions, and BP states. Existing methods often rely on handcrafted features or direct regression from short segments. These designs may not fully capture subject-specific temporal changes. This paper proposes DeltaMamba, a differential pretrained state-space framework for PPG-based cuffless BP estimation. DeltaMamba first adapts a pretrained optical physiological signal encoder to extract task-relevant PPG representations. It then builds subject-specific differential features by subtracting an individual baseline representation. This step reduces static inter-subject bias and highlights BP-related temporal variation. Finally, a Mamba-based selective state-space regressor models the differential sequence to estimate systolic BP (SBP) and diastolic BP (DBP). Experiments on the public MIMIC-BP benchmark show that DeltaMamba produces a more concentrated error distribution than a feature-based differential baseline. It also outperforms several alternative temporal backbones with lower computational cost. These results suggest that pretrained PPG representations and selective state-space modeling are effective for continuous cuffless BP estimation.