The stability metric Phi discriminated arrhythmia from normal sinus rhythm with an AUC of 0.9015, ranking behind the cardiac-specific RMSSD metric (AUC 0.9706) in a head-to-head comparison.
Does the unsupervised stability metric Phi accurately detect cardiac arrhythmias and physiological instability compared to standard heart-rate-variability metrics?
A novel, unsupervised cross-domain stability metric (Phi) demonstrated high accuracy for detecting cardiac arrhythmias, though it was outperformed by the cardiac-specific RMSSD metric.
Absolute Event Rate: 0.9015% vs 0.9706%
Physiological monitoring methods for detecting rhythm instability typically rely on signal-specific feature pipelines, supervised classifiers trained on labeled examples, or modality-specific scoring rules that do not transfer across measurement domains. This paper evaluates a stability metric requiring no supervised model training, Phi = I × rho - alpha × S, computed from physiological time-series data using domain-appropriate observable mappings. On the MIT-BIH Arrhythmia Database, under per-record threshold calibration, Phi discriminated arrhythmia from normal sinus rhythm at AUC 0.9148 across 1,022 evaluable 60-second windows from 34 records, with a shuffle audit AUC of 0.5003 confirming the signal was not an artifact of label leakage. On the MIT-BIH Atrial Fibrillation Database, Phi produced AUC 0.5556 across 11,988 windows, a result interpreted as a scope boundary supporting the structural reading that Phi tracks transitions rather than classification between stable rhythm modes. Resolution scaling on the same cardiac dataset produced AUC 0.8775 at 30-second windows and AUC 0.6376 at 5-second windows, indicating graceful degradation rather than abrupt collapse. A head-to-head comparison against 16 standard heart-rate-variability metrics on 811 matched windows placed Phi at AUC 0.9015, sixth among 16 metrics, behind cardiac-specific RMSSD (0.9706) by 0.0691 AUC. Exploratory neural validation on the CHB-MIT Scalp EEG Database produced weak and direction-dependent window-level results across four tested patients. A 3-consecutive-hit audit on chb01 falsified an earlier reported 71% event-level seizure detection result by reducing detection to 0 of 7 while leaving window-level AUC unchanged at 0.5654, consistent with a chance-rate calculation that gave approximately 78% probability of at least one chance hit per seizure under the original single-hit criterion. A K-of-N event-detection rule (3 hits in any 10 windows) increased single-patient event sensitivity to 67% (4 of 6 seizures with sufficient coverage) at 0.74 false alarms per hour. Results are bounded to the tested datasets, observable mappings, and configurations. Generalization beyond these specific choices, including to other physiological modalities, to multi-patient validated seizure prediction, or to actual wearable hardware deployment, is not established by the evidence reported here. The methods described in this paper are the subject of U.S. Provisional Patent Application No. 63/978,132 (filed February 9, 2026). No license to implement or commercialize the described methods is granted by this publication. All rights reserved.
Shawn Barnicle (Sat,) conducted a other in Cardiac Arrhythmia and Physiological Instability. Stability metric Phi vs. 16 standard heart-rate-variability metrics (including RMSSD) was evaluated on Discrimination of arrhythmia from normal sinus rhythm (AUC). The stability metric Phi discriminated arrhythmia from normal sinus rhythm with an AUC of 0.9015, ranking behind the cardiac-specific RMSSD metric (AUC 0.9706) in a head-to-head comparison.