ARTA: Graph Combinatorial Optimization for Robust Interbeat Interval Estimation from Wearable Electrocardiogram Signals

Abstract

Accurate heart rate (HR) and interbeat interval (IBI) estimation from wearable electrocardiogram (ECG) is essential for continuous physiological monitoring, but motion artifacts can create false QRS-like peaks and obscure true R-peaks. Conventional detectors and local interval-repair rules process detections independently and may become unreliable when several consecutive beats are corrupted. We present ARTA (Artifact-Robust Temporal Alignment), a graph-based combinatorial framework that estimates the heartbeat sequence through constrained temporal optimization. ARTA combines candidates from multiple ECG detectors, represents them as nodes in a directed acyclic graph and connects candidates through physiologically plausible RR intervals. Candidate confidence uses detector agreement and local ECG quality, while history-dependent transition costs capture recent IBI consistency and a soft heart-rate prior. Beam-search dynamic programming selects a coherent low-cost path across the recording, suppressing false detections and recovering missed beats when plausible auxiliary candidates are available. On the IEEE Signal Processing Cup exercise dataset with electrode-motion artifacts from -2 to 12 dB, ARTA achieved an HR mean absolute error of 1.79 beats per minute, 54% lower than the strongest single-detector baseline, with 92.38% of windows within +/-5 beats per minute, an R-peak F1 score of 0.95, and an IBI mean absolute error of 6.52 ms. Cross-dataset evaluation on the European ST-T Database further assessed transfer to long clinical ambulatory recordings and a label-free reliability analysis examined conditional performance without using reference annotations for segment selection. ARTA is single-channel, training-free, interpretable and computationally lightweight, operating approximately 27 times faster than real time with peak memory below 2.1 MB in a single-thread CPU simulation.

Shovito Barua Soumma

Graduate Research Associate
PhD Candidate

Currently I am working on building and optimizing deep learning models for wearable sensors data.