The physiological logic behind AVCI is elegant. In a healthy heart, breathing drives two parallel processes at the same frequency. Through the vagus nerve, each breath modulates the timing of heartbeats, producing respiratory sinus arrhythmia, the well-known fluctuation of RR intervals at the breathing frequency. At the same time, each breath alters intrathoracic pressure and venous return, which changes ventricular filling and, through the Frank–Starling mechanism, modifies the QRS area, a morphological signature of ventricular depolarization. Because both signals oscillate in the same respiratory band, they should show high spectral coherence in healthy hearts. Heart failure disrupts this harmony: sympathetic overdrive and vagal withdrawal flatten the respiratory component of RR variability, while ventricular remodeling makes QRS morphology less responsive to autonomic input. Premature ventricular contractions are even more destructive, inserting large random perturbations into QRS morphology that have no counterpart in autonomic modulation, directly destroying coherence at respiratory frequencies.

Computationally, AVCI is built from two beat-indexed signals. For each detected R-peak, the QRS area is computed as the trapezoidal integral of the absolute filtered ECG over a fixed 120-millisecond window centered on the peak, a morphology feature that requires no delineation of the Q, S, or T waves. The RR interval is simply the time between successive R-peaks, filtered to a physiological range of 400 to 1,500 milliseconds to exclude false detections. From these features, the researchers constructed an autonomic variability signal, defined as the absolute beat-to-beat change in RR intervals, and a ventricular morphology variability signal, defined as the absolute beat-to-beat change in QRS area. Both signals were z-score normalized, resampled onto a uniform 1-hertz time grid by linear interpolation, and then analyzed with Welch’s method for magnitude-squared spectral coherence. The coherence values were integrated over the high-frequency band from 0.15 to 0.40 hertz, which spans normal breathing rates of roughly 9 to 24 breaths per minute and captures respiratory sinus arrhythmia. Subject-level AVCI values were averaged over sliding 5-minute windows with 1-minute steps, yielding a single interpretable scalar per recording.

The evaluation drew on 88 recordings from four PhysioNet databases: 18 long-duration recordings from the MIT-BIH Normal Sinus Rhythm Database sampled at 128 hertz, 15 recordings from the BIDMC Congestive Heart Failure Database of patients with advanced NYHA Class III–IV failure sampled at 250 hertz, 48 half-hour recordings from the MIT-BIH Arrhythmia Database sampled at 360 hertz, and 7 recordings from the MIT-BIH Long-Term ECG Database reserved for preliminary independent validation. Because these databases differ in sampling rate and recording length, the researchers examined AVCI under each database’s native conditions and after harmonizing all recordings to a common 128-hertz, 30-minute standard, with additional upsampling to 360 hertz to separate sampling-rate effects from duration effects. R-peak detection, validated against expert annotations with a ±50-millisecond tolerance, achieved an F1-score of 0.892, rising to 0.910 when two extremely noisy arrhythmia recordings were excluded.

The results under native conditions were striking. Healthy subjects showed a mean AVCI of 0.663, compared with 0.526 for heart failure patients, 0.506 for PVC-dominant arrhythmia recordings, and 0.360 for the two PAC recordings, the lowest values in the study. Statistical testing confirmed these differences: the Kruskal–Wallis test yielded a highly significant result across groups, and pairwise comparisons with Bonferroni correction showed that healthy rhythm differed significantly from heart failure, PVC-dominant recordings, and mixed arrhythmia recordings, with large Cliff’s delta effect sizes of 0.689, 0.767, and 0.675 respectively. The diagnostic comparison against traditional heart rate variability metrics was particularly revealing. AVCI achieved an AUC of 0.844 for distinguishing normal rhythm from heart failure, while SDNN managed only 0.681, RMSSD 0.567, and mean heart rate 0.374, below the chance level. A Random Forest classifier combining AVCI with the three HRV metrics attributed 45.8 percent of overall feature importance to AVCI, more than the combined contribution of all three conventional measures.

Machine learning reinforced the picture. Eight classifiers, including logistic regression, decision trees, random forests, support vector machines, gradient boosting, K-nearest neighbors, Gaussian Naïve Bayes, and XGBoost, were trained on balanced datasets under native conditions using repeated stratified five-fold cross-validation. K-nearest neighbors achieved 79.5 percent accuracy with a ROC-AUC of 85.7 percent, while support vector machines reached 79.0 percent accuracy with a ROC-AUC of 83.0 percent. Internal cross-validation of the AVCI decision threshold showed that the best cut-off remained stable across folds, ranging from 0.543 to 0.557, closely matching the value between the healthy and heart failure group averages. Held-out performance from 200 random train-test splits averaged an AUC of 0.846, essentially identical to the full-sample estimate, indicating that the index’s discrimination is not an artifact of fitting a particular sample.

The preliminary independent validation added an external check. Applying the identical, unchanged pipeline to six recordings from the MIT-BIH Long-Term ECG Database, a cohort with ventricular ectopy sampled at the same 128 hertz as the healthy database, produced a mean AVCI of 0.476, clearly in the unhealthy range and significantly lower than the healthy average, with a large Cliff’s delta of 0.870. The values closely resembled those of the heart failure group, consistent with the expected reduction in autonomic–ventricular coupling in diseased hearts. Although the independent cohort was small, the agreement across databases supported the index’s robustness as a marker of healthy versus unhealthy cardiac status.

Then came the caveat that gives the study its title. After harmonizing all recordings to a common 128-hertz, 30-minute standard, the discrimination collapsed. The AUC for distinguishing normal rhythm from heart failure fell to 0.647, and the AUC against all disease recordings fell to 0.619, comparable to SDNN’s performance under the same conditions. None of the pairwise group differences remained statistically significant after correction, and the machine learning classifiers produced only average results, with accuracies ranging from 57.9 to 66.5 percent on balanced harmonized data. A two-tier duration-stratified analysis clarified the mechanism: Tier 1, comparing long-duration healthy and heart failure recordings of roughly 19 hours, preserved a significant difference with an AUC of 0.844, while Tier 2, a 30-minute snapshot analysis, did not reach significance. QRS-window sensitivity testing confirmed that the collapse was not an artifact of window selection, with discrimination remaining low across window widths from 80 to 180 milliseconds under harmonized conditions, in contrast to the stable native-condition performance of 0.826 to 0.859. Recording duration, not morphology parameters, drives the index’s discriminative strength.

The authors are transparent about the interpretation. AVCI’s advantage over traditional HRV appears under native conditions but converges with conventional measures after acquisition standardization, meaning the index does not retain an edge when recordings are harmonized. The physiological explanation of AVCI as autonomic–ventricular coupling remains a hypothesis, since the retrospective databases contained no simultaneous measurements of respiration, autonomic activity, or ventricular mechanics. The AF and PAC subgroups consisted of only two recordings each and were reported descriptively. The arrhythmia recordings, at roughly 30 minutes, were far shorter than the 24-hour healthy and 20-hour heart failure recordings, which may partly explain the lack of significance between healthy rhythm and arrhythmia-normal recordings after correction. Larger, prospectively defined populations with standardized acquisition protocols will be needed before clinical use.

Still, the practical appeal is considerable. AVCI requires only a single ECG lead, no additional physiological signals, no multi-lead recordings, and no manual waveform marking, and it can be computed from recordings sampled at moderately low rates such as 128 hertz. The researchers report that analyzing a 24-hour ECG takes less than two minutes on a standard computer, making the method compatible with Holter monitors, wearable ECG devices, and mobile cardiac telemetry. Unlike deep learning ECG models that offer accuracy but little physiological insight, AVCI yields a single value grounded directly in the signal and its underlying mechanism, an attribute the authors argue is valuable for clinical practice, regulatory evaluation, and routine cardiovascular risk assessment. The analysis code is publicly available on GitHub, and the framework’s combination of coherence-based biomarkers with interpretable machine learning offers a roadmap toward cardiovascular diagnostics that are both powerful and explainable. The crucial next step, the authors conclude, is to acquire uniform, long-duration single-lead ECG recordings from healthy and diseased individuals, ideally with simultaneous respiration and independent autonomic and ventricular-mechanical measurements, to properly test the proposed physiological interpretation and isolate the computational association measured by AVCI from differences in acquisition and population.

Subject of Research: The autonomic-ventricular coherence index: a novel single-lead ECG method with acquisition-dependent discrimination

Article Title: The autonomic-ventricular coherence index: a novel single-lead ECG method with acquisition-dependent discrimination

Article References: R, J., & G, C. S. (2026). The autonomic-ventricular coherence index: a novel single-lead ECG method with acquisition-dependent discrimination. Frontiers in Digital Health, 8, Article 1964571. https://doi.org/10.3389/fdgth.2026.1964571

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1964571

Keywords: autonomic-ventricular, coherence, index, novel, single-lead, method, acquisition-dependent, discrimination, scientific research