BiMa-ECGSR: 12-lead ECG reconstruction from reduced lead sets

Runtime: ZeroGPU (a GPU is attached per request); device cuda; Mamba backend: mamba_ssm CUDA kernels if they pass a one-window self-test against the pure-PyTorch Mamba (|Δ| < 0.001), else the pure-PyTorch Mamba on CUDA; mamba_ssm importable: yes; 14 model instances loaded in 37.1 s.

ECG source
Example record

Examples (fixed rule): within each label group, the record at the 75th percentile of BiMa-SR per-record PCC for the Telemetry configuration (I+II+V1); PTB-XL test-fold groups NORM, MI, STTC, CD, HYP (single superclass), CLBBB, CRBBB, AFIB, LVH (single panel label), plus one record each from Georgia and Chapman-Shaoxing/Ningbo. Data: PTB-XL (Wagner et al. 2020), Georgia (PhysioNet/CinC Challenge 2020), Chapman-Shaoxing/Ningbo (Zheng et al. 2020), via PhysioNet; all CC BY 4.0.

Measured input leads (use case)

Model (DiBiMa-SR: Monitor / II only)

Per-lead PCC / MAE (z units); summary rows: pooled median over windows x leads, as in the paper