Ensemble AI decodes intracardiac heart sounds to distinguish ventricular and atrial fibrillation

Inside the Beating Heart, AI Learns to Hear the Difference Between Two Deadly Rhythms

Deep inside the chest, a heart in trouble makes a very particular kind of noise — and researchers in China have now taught an artificial intelligence to hear the difference between two of the deadliest ways that noise can go wrong. In a study published on 29 August 2026 in Medical & Biological Engineering & Computing, engineers at Dalian University of Technology and cardiologists at the First Affiliated Hospital of Dalian Medical University unveiled a framework that analyzes sounds recorded from within the heart itself and separates atrial fibrillation from ventricular fibrillation with 96.7 percent accuracy. The achievement matters because these two rhythms sit at opposite ends of the cardiac emergency spectrum: atrial fibrillation is the most common sustained arrhythmia in clinical practice and a major cause of stroke, while ventricular fibrillation is the chaotic electrical storm that shuts down the heart’s pumping action and can kill within minutes. The system paired that accuracy with a precision of 96.5 percent, a recall of 96.8 percent and an F1 score of 96.6 percent — performance figures the team reports make the framework a dependable candidate for continuous monitoring of abnormal intracardiac heart sounds.

The two conditions the algorithm was trained to separate are as different in mechanism as they are in urgency. In atrial fibrillation, the heart’s upper chambers — the atria — abandon their coordinated contraction and quiver chaotically, allowing blood to pool and clot; those clots can travel to the brain and trigger ischemic stroke, which is why international registries that have followed tens of thousands of emergency-department patients across dozens of countries consistently identify the arrhythmia as a global health burden. Ventricular fibrillation is even more unforgiving: the lower chambers — the ventricles — twitch uselessly instead of contracting, cardiac output collapses, and sudden cardiac death follows unless a defibrillating shock is delivered almost immediately. Clinicians ordinarily track both conditions through the electrocardiogram, which records the heart’s electrical activity from the skin. But electrical signals can be obscured by noise, motion and electrode problems, and they do not directly capture the mechanical fingerprint of the beating heart. Sound offers a complementary window: every heartbeat generates turbulence as blood surges through valves and chambers, and the texture of that turbulence changes when rhythm degenerates.

What sets the new study apart is where its data come from. Rather than listening through a stethoscope pressed against the chest — a technique that captures heart sounds diluted by skin, fat, muscle, lung tissue and ambient noise — the team worked with a high-fidelity sonocardiogram dataset in which acoustic signals were acquired from inside the heart. The recordings were obtained in porcine experiments approved by the Animal Care and Use Committee of Dalian Medical University, with all animal handling, anesthesia and post-procedural care carried out under the approved protocol. Pigs are a mainstay of cardiovascular research because the size, anatomy and electrical behavior of their hearts closely resemble those of humans, making them a demanding and clinically meaningful proving ground for any monitoring technology. Intracardiac acoustics remain far less explored than the surface phonocardiogram, yet they carry an unusually clean signal: recorded at the source, they preserve the fine turbulence produced by valvular motion and blood flow that conventional chest-wall recordings smear or lose entirely.

Before any learning could happen, the raw recordings were shaped by two classic signal-processing steps. First, the team applied a window function, a mathematical taper that smoothly fades the beginning and end of each analyzed segment toward zero. Without it, cutting a continuous recording into short frames creates sharp discontinuities that smear energy across the frequency spectrum — a phenomenon known as spectral leakage — and bury the delicate acoustic signatures of abnormal rhythm. Second, the signals passed through a pre-emphasis filter, a simple high-pass boost that amplifies higher frequencies relative to lower ones. Heart sounds concentrate much of their energy in low-frequency rumble, yet many of the diagnostically valuable details of turbulent flow sit in the higher bands, where they are easily masked. By flattening the spectrum before analysis, pre-emphasis ensures that the algorithms downstream pay as much attention to those faint high-frequency clues as to the thunderous low-frequency components that dominate the raw waveform.

The heart of the method lies in its refusal to trust any single way of describing a sound. The researchers extracted four complementary families of features from each recording. Mel-frequency cepstral coefficients, a representation borrowed from speech recognition, summarize the spectral envelope of the signal on a frequency scale that mimics human pitch perception; in this setting they encode the timbre of valve closures and fibrillatory murmurs. Envelope autocorrelation measures how strongly the loudness of the signal repeats from one moment to the next, providing a direct mathematical probe of rhythmic structure — a property whose signature differs sharply between the irregular but still cyclic activity of atrial fibrillation and the nearly patternless chaos of ventricular fibrillation. The Hilbert-Huang transform, an adaptive technique built on empirical mode decomposition, breaks a signal into its natural oscillatory components and tracks how their instantaneous frequency drifts over time, making it well suited to sounds that are nonlinear and non-stationary by nature. Finally, the wavelet scattering transform pushes the signal through a deep cascade of wavelet filters and averaging operators, yielding descriptors that are stable to small time shifts and sensitive to multi-scale texture.

Each family of features describes the same heartbeat from a different angle, and simply concatenating them produces a high-dimensional vector in which many dimensions are redundant or irrelevant. To tame that complexity, the team turned to the maximum relevance minimum redundancy algorithm, a feature-selection strategy that scores candidate variables on two competing criteria: how strongly each one is associated with the diagnostic label, and how little information it duplicates in features already chosen. By iteratively retaining variables that are maximally relevant to the fibrillation type while minimally redundant with one another, the algorithm distills a compact, information-dense input for the machine-learning stage. The benefits are practical as well as statistical. Smaller feature sets train faster, demand less memory, and are less prone to overfitting — the failure mode in which a classifier memorizes quirks of its training data rather than learning the underlying physics of the disease. In a monitoring system that might one day run continuously, that efficiency is not a luxury; it is a design requirement.

On top of the fused features sits the framework’s second pillar: a voting ensemble that combines three fundamentally different base classifiers — K-nearest neighbors, support vector machines and artificial neural networks. K-nearest neighbors classifies a new recording by looking at the labels of the most similar examples in its training set, an approach that adapts naturally to irregular, local structure in the data. Support vector machines draw the boundary between classes by maximizing the margin of separation, a principle that tends to generalize well when the two categories overlap. Artificial neural networks, meanwhile, learn layered nonlinear transformations that can capture interactions among features that simpler models miss. Each algorithm was allowed to cast a vote on whether a given recording represented atrial or ventricular fibrillation, and the majority verdict set the final decision. Because the three models stumble on different kinds of samples, their errors largely cancel out in the vote. The result was an overall accuracy of 96.7 percent, with precision of 96.5 percent, recall of 96.8 percent and an F1 score of 96.6 percent — a balanced profile indicating that the system is equally reluctant to raise false alarms and to miss true cases.

The designers argue that the ensemble’s strength lies precisely in the complementarity of its members: by leveraging the distinct strengths of instance-based, margin-based and network-based learning, the framework avoids depending on any one mathematical worldview. That robustness is exactly what a clinical alarm system needs. A missed episode of ventricular fibrillation can be fatal, while a spurious diagnosis of atrial fibrillation can set off a cascade of unnecessary anticoagulation and its associated bleeding risks, so both kinds of error carry a price. The researchers position the framework as a reliable tool for abnormal intracardiac heart sound monitoring systems — technology that could eventually sit alongside electrical monitoring in intensive care units, cardiac catheterization laboratories and implantable or catheter-based devices. Because the approach processes sound rather than electrical potentials, it could in principle keep working in situations where electrocardiogram signals degrade, and it adds an entirely independent physical observable to the clinician’s picture of what the heart is doing.

Like any early-stage result, the study comes with caveats that its authors acknowledge. The sonocardiogram dataset is not publicly available, owing to institutional data-sharing restrictions, though the corresponding author, Yunlong Xia, can provide it upon reasonable request — a policy that will make independent benchmarking harder in the short term. The experiments were performed in a porcine model, so performance in human patients, whose chest and cardiac anatomy vary far more widely, remains to be demonstrated. And because the analysis focuses on distinguishing the two fibrillations from each other, further work will be needed before the method can triage the full spectrum of arrhythmias that a monitoring system might encounter. Even so, the study lands at a moment when machine learning is rapidly colonizing cardiac acoustics, with prior efforts demonstrating that heart sounds alone can flag abnormality, and the new results suggest that recordings taken closer to the source — combined with disciplined feature engineering and ensembles — can push that capability toward the level of accuracy previously associated with far more invasive electrical measurements.

The work was supported by China’s National Key Research and Development Program under grants 2022YFC2405002 and 2022YFC3601301, a signal of how seriously national research programs now treat intelligent cardiac monitoring. The research team spans two institutions — Yiming Xin, Zhe Chen and Fuliang Yin of Dalian University of Technology’s School of Information and Communication Engineering, together with Rongfeng Zhang, Yiheng Yang and Yunlong Xia of the First Affiliated Hospital of Dalian Medical University — an alliance of signal-processing expertise and bedside cardiology that mirrors the multidisciplinary nature of the problem itself. If subsequent clinical studies confirm what the porcine recordings show, the implications extend beyond the catheterization laboratory: stethoscopes, pacemakers and monitoring catheters of the future could carry an algorithm that never stops listening, never gets tired and never mistakes one deadly rhythm for another. When the heart loses its rhythm, the new research suggests, it also loses its tune — and a machine has now been taught to hear exactly which tune went missing.