FiLMamba: A FiLM-Conditioned Mamba Architecture with an Energy-Efficient FPGA Accelerator for Generalizable Cuffless Blood Pressure Estimation

Cuffless blood pressure (BP) estimation from photoplethysmography (PPG) is limited by a morphology-BP ambiguity that similar waveforms map to different pressures across subjects, causing population-trained models to collapse into mean regression. We propose FiLMamba, a selective state-space model conditioned via Feature-wise Linear Modulation (FiLM) on a 4-D hemodynamic state vector, and Cross-Domain Hemodynamic Alignment (CDHA), which orthogonally decomposes residual cross-corpus shift into unlabeled batch-normalization recalibration and additive bias correction per calibration unit. On MIMIC-BP across 1,524 patients, FiLMamba achieves SBP/DBP standard deviation of error (SDE) of 6.45/4.96mmHg with >2.5× improvement over unconditioned baselines, satisfying Association for the Advancement of Medical Instrumentation (AAMI) compliance. Under training-free transfer, it further reaches 4.29/3.24mmHg on UCI-BP, the first AAMI-compliant training-free cross-dataset result in cuffless BP estimation. A dedicated FPGA accelerator specializes this stack on Zynq UltraScale+ through Heterogeneous Engine Reuse Orchestration (HERO) and Persistent Recurrent In-SRAM Mamba (PRISM), operating at 0.514W dynamic power and 150MHz with 8.23 μJ per sample, 15.3-41.1× below evaluated CPU/GPU platforms, and an energy-delay product of 0.046 J·s that is 1.87-14.8× below every general-purpose baseline.