AI Detects Heart Problems from Sound Alone

Heart rhythm disorders affect millions worldwide, yet many go undiagnosed until serious complications occur. A new artificial intelligence system can now identify these dangerous irregularities simply by listening to heart sounds, achieving near-perfect accuracy that could transform how we monitor cardiac health.

The research team developed a specialized AI that detects arrhythmias—abnormal heart rhythms—from audio recordings of heartbeats. Their system correctly identified 99.42% of irregular heartbeats in tests, significantly outperforming existing methods. This breakthrough means potentially life-threatening heart conditions could be spotted earlier through simple audio analysis rather than complex medical tests.

The approach combines two AI techniques: convolutional neural networks (CNNs) that analyze visual representations of heart sounds, and long short-term memory (LSTM) networks that track patterns over time. The key innovation comes from adapting control theory mathematics—specifically H-infinity filtering—to help the AI maintain accuracy even with noisy or limited data. Think of it as giving the AI a built-in correction mechanism that adjusts for imperfections in real-world recordings.

Researchers trained and tested their system using the PhysioNet CinC Challenge dataset, containing approximately 6,000 heart sound recordings. The AI converts audio into spectrograms—visual representations of sound frequencies—then analyzes these images for patterns indicating healthy versus irregular rhythms. The system achieved 99.23% sensitivity (correctly identifying true problems) and 99.49% specificity (correctly ruling out non-problems), outperforming all benchmark models including ResNet-50 (88.94% accuracy) and MobileNetV3-Large (95.23% accuracy).

This technology matters because heart disease remains the leading global cause of death, claiming nearly 18 million lives annually. Current diagnosis often requires expensive equipment and specialist interpretation. The new method could enable affordable screening through smartphones or simple recording devices, particularly valuable in remote areas or for continuous monitoring of at-risk patients. The system's design allows potential deployment on mobile devices, supporting low-cost, scalable heart health screening.

The research acknowledges limitations, including working with a dataset where 87% of recordings represented normal heart sounds versus only 13% abnormal cases. While the team developed special techniques to handle this imbalance, real-world performance across diverse populations and recording conditions requires further validation. The system also currently focuses on identifying the presence of arrhythmias rather than distinguishing between specific types of rhythm disorders.