Rot-IIR-SSM: Provably Stable and Pole-Interpretable IIR State-Space Model for Streaming ECG Myocardial Infarction Detection.

Abstract

Many modern state-space models (SSMs) for ECG analysis lack simultaneous stability guarantees, spectral interpretability, and hardware-efficient streaming inference. We introduce Rot-IIR-SSM, a rotational IIR state-space model in which each channel is parameterized by a stable second order complex pole pair (ρ,θ). Bounded polar reparameterization guarantees BIBO stability by construction; an FFT-based realization enables O(T logT) parallel training with explicit pole domain interpretability; and the learned parameters deploy as biquad-filter recursions for O(1) streaming inference requiring only ≈25KB of state memory. On 12-lead ECG MI detection (PTB-XL five-fold CV; PTB-DB external test), Rot-IIR SSM achieves the highest external F1 without post-processing (0.8306) and competitive internal F1 at the clinical operating point (0.8360 vs. CNN 0.8508), with near-equivalent threshold independent AUROC (0.9251 vs. 0.9337). A supplemental atrial fibrillation experiment provides evidence of applicability to a second ECG task (F1 = 0.8939). Five-fold instance-level frequency masking (one held-out validation case per fold, filtered by maximum diagnostic annotation confidence and ranked by model prediction probability, independent of spectral content) identifies the 10-20Hz (QRS morphology) band as the consistent dominant driver (ΔP = +0.175 ± 0.141), with negligible contribution from frequencies above 40Hz. Ablation studies suggest that second-order state dimensionality plays a more important role than frequency selectivity in the observed performance gains. Rot-IIR-SSM simultaneously provides pole-domain interpretability, provable BIBO stability, and O(1) streaming-a combination not jointly exhibited by the evaluated baselines in this study. This combination may be advantageous in resource-constrained, interpretability-demanding clinical ECG deployment settings.