Adaptive Autocorrelation Based Heart Rate Estimation From Single-Axis Seismocardiogram: A Comprehensive Benchmark Across Six Diverse Datasets.

We introduce AACFD (Adaptive Autocorrelation Function Detector), a lightweight, fully automatic pipeline for estimating window-averaged heart rate (HR) from a single-axis seismocardiogram (SCG) without ECG calibration or machine learning. AACFD combines two periodicity detectors-(i) a YIN-style difference-function analysis of an adaptive SCG envelope and (ii) a short-window autocorrelation branch-operating on 3 s segments. The two branches are fused by feature-aware weighting based on signal-quality indices and multi-scale features; and a Hampel filter plus temporal consistency checks remove outlier windows. The algorithm is efficient enough for real-time implementation on commodity hardware. AACFD was validated on 535 subjects spanning six heterogeneous datasets (sampling rates 200 Hz-5 kHz): laboratory mechanocardiograms (MCG), controlled-breathing SCG (CEBS), motion-rich multichannel SCG (MC-SCG), valvular heart-disease clinics (VHD), invasive right-heart catheterization (RHC), and smartphone recordings from 300 atrial-fibrillation/sinus-rhythm subjects. For each recording we analyzed non-overlapping windows of 10, 20, 30, and 60 s and compared the SCG-derived mean HR with ECG-derived (or telemetry) reference values. After ECG- and SCG-based quality control and temporal refinement, the mean per-subject mean absolute error (MAE) on 30-s windows was 0.99 bpm in MCG, 0.72 bpm in CEBS, 4.38 bpm in VHD, 1.34 bpm in MC-SCG after removal of recordings with corrupted ECG, 3.58 bpm in RHC, and 7.67 bpm in the smartphone cohort. Across all datasets, more than 90% of non-overlapping 10-60 s windows passed ECG- and SCG-based quality control, so the reported errors reflect nearly the full usable recording duration rather than a few selected clean segments. This cross-dataset benchmark indicates that a short-window hybrid YIN and autocorrelation routine, guarded by simple robustness checks, attains sub-bpm accuracy on resting datasets and clinically acceptable accuracy on several pathological cohorts using only a single accelerometer axis. Importantly, AACFD is deliberately designed and evaluated for reliable window-averaged HR estimation on 10-60 s time scales rather than instantaneous beat-to-beat HR; deriving robust heart-rate-variability indices from SCG remains an important direction for future work.