From 805 ms to 23 ms: Accelerating State-Space Models for Real-Time ICU Monitoring with Fused Triton Kernels

Real-time ICU early-warning systems operate under low-latency constraints, with sub-50 ms targets enabling timely clinical intervention. However, irregular sampling and missing values exceeding 30\% in physiological time series force deep learning pipelines into sequential preprocessing routines that consume over 85\% of total wall-clock time. We introduce Triton-accelerated state-space models (SSMs) with a time-aware formulation that fuses interpolation and inference into a single GPU kernel, eliminating this bottleneck. Our system reduces end-to-end inference latency by $35.7\times$ (from 805 ms to 23 ms) compared to a PyTorch SSM implementation, achieves $2.5\times$ faster training than GRU-D baselines, and scales to 10,316 samples per second. Across 5 seeds on PhysioNet Challenge 2012, the Triton SSM improves test AUROC over GRU-D by $+0.037$ (paired bootstrap 95\% CI $[0.018, 0.058]$, $p<0.01$); the same kernel ports to MIMIC-III 25-task phenotyping with a $+0.024$ macro-AUROC improvement. The implementation is hardware-portable across NVIDIA and AMD GPUs.