Low-latency HRV analysis from ultra-short ECG windows using a modular deep-learning framework

We present a universal modular deep-learning framework and demonstrate its application to low-latency, streaming-compatible heart rate variability (HRV) analysis using RMSSD as an exemplar metric. A convolutional autoencoder is first pretrained and then reused as a frozen encoder that maps raw ECG windows to a compact latent sequence. Task-specific heads, each comprising a BiLSTM adapter, a shallow Conv1D refinement, and temporal attention pooling operate on this shared representation. A discriminator head screens low-quality windows, while a regression head estimates RMSSD; a gated inference block routes outputs so RMSSD is produced only when the discriminator exceeds a threshold, replicating a robust "mask-then-estimate" pipeline in a single deployable graph. Using LUDB and PTB-XL with segmentation-assisted peak extraction for PTB-XL, plus an out-of-distribution Apple Watch subset, we enforce rigorous quality assurance to derive validity labels and RMSSD targets. Compared to two strong classical baselines (HeartPy and NeuroKit2), our discriminator improves combined-set accuracy to 92.12% (vs. 80.54% / 85.58%) with F1 of 95.43% (vs. 88.82% / 91.99%). On RMSSD estimation, the proposed model reduces combined MAE to 10.56 ms (from 45.12 ms / 27.93 ms) and sharply curtails tail errors (P95: 47.00 ms vs. 313.35 ms / 167.84 ms), indicating substantially improved robustness under pathological and noisy ECG. On a small out-of-distribution Apple Watch subset used as a sanity-check for acquisition shift, where the model attains the lowest MAE (7.57 ms vs. 13.96 ms / 9.61 ms) under a selective gating regime. The end-to-end model is compact (2.62 M parameters; 10.07 MB on disk) and real-time capable, achieving 15.0 ms mean latency at batch size 1 (66.5 windows/s) and scaling to ∼4.49k windows/s at batch size 1024 on a single consumer-grade GPU.