Multimodal AI estimates hemodynamic echocardiographic indicators of structural heart disease from heart sounds and pulse waves
This study demonstrates that a multimodal AI model combining heart sounds, pulse waves, and demographics can generate triage-level estimates of key echocardiographic hemodynamic indicators from a single 15-second recording, achieving high sensitivity for ruling out structural heart disease despite predictions being compressed toward the cohort mean and limited by single-center data.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
For decades, the stethoscope has been the universal symbol of the doctor's first look at a patient's heart. It is a simple tool that listens to the rhythm of life, the closing of valves, and the rush of blood. When a doctor hears a murmur—a whooshing sound that shouldn't be there—they know something is wrong. But the stethoscope has always been a qualitative instrument; it tells a physician that a problem exists, but it cannot tell them exactly how severe that problem is. To get those precise numbers, a patient must undergo an echocardiogram, an ultrasound of the heart that requires expensive machinery and a highly trained specialist to operate. This creates a gap in modern medicine: the tool available at the very first moment of contact cannot measure the heart's pressure and flow, while the tool that can measure them is often unavailable until much later. As a result, many cases of heart valve disease go undetected for years, leaving patients at risk.
A team of researchers in South Korea has attempted to bridge this gap by teaching a computer to listen to the heart in a new way. They did not invent a new machine or add new sensors to the stethoscope. Instead, they used a single device that records both the sound of the heart and the pulse wave from a fingertip simultaneously. By feeding these two streams of data into a sophisticated artificial intelligence model, the researchers asked a bold question: can the computer extract the exact same numbers that a cardiologist usually gets from an ultrasound, using only fifteen seconds of sound and a pulse reading? The answer is a qualified yes. The system can now estimate key hemodynamic indicators—such as the speed of blood flowing through a valve or the pressure in the lung arteries—with a level of accuracy that makes it a powerful tool for triage, or sorting patients who need immediate specialist care from those who do not.
The study involved over 1,300 patients who came to a hospital for a standard heart ultrasound. Before the ultrasound began, each patient wore a specialized electronic stethoscope on their chest for fifteen seconds while a sensor on their finger recorded their pulse. The researchers then used these recordings to train an artificial intelligence model to predict the specific measurements that would appear on the ultrasound report. The model was designed to look for patterns in the sound and the pulse that correspond to the mechanical work of the heart. It learned to estimate values like the peak velocity of blood moving through the aortic valve, the pressure in the pulmonary artery, and the volume of blood the heart pumps with each beat.
The results showed that the computer could indeed make these estimates. Across a wide range of heart conditions, the model's predictions correlated reasonably well with the actual ultrasound measurements. For some indicators, such as the speed of blood flow through the aortic valve and the pressure in the pulmonary artery, the agreement was quite strong. The system was particularly good at detecting regurgitation, a condition where blood leaks backward through a valve, because the sound of that leak carries specific information that demographics like age and sex cannot provide. However, the system was less accurate at estimating the size of the heart chambers or the thickness of the heart walls. This makes sense physically: the sound of the heart is generated by the movement of blood, so the computer is excellent at measuring things related to flow and pressure, but it cannot "hear" the static structure of the heart muscle itself.
Crucially, the researchers found that the computer's estimates are not perfect measurements to replace an ultrasound. The predictions tend to be compressed toward the average of the group, meaning the system is less likely to guess extreme high or low values. Because of this, the output is best used as a screening tool rather than a final diagnosis. The team demonstrated that by using a simple rule based on these estimates, they could identify nearly all patients who had dangerous levels of valve disease or high blood pressure in the lungs. In a test scenario, this approach would have caught 95% of the patients who truly needed an ultrasound, while sparing about one in four patients from undergoing the test unnecessarily. This suggests the technology could serve as a highly effective filter in clinics or community settings, ensuring that those with serious conditions are seen by specialists quickly.
The study also revealed the distinct roles of the two signals the device collected. When the researchers tested the model by removing the heart sound data, it lost its ability to predict flow-related issues like valve leaks and pressure gradients. When they removed the pulse wave data, the model struggled to estimate the size of the heart chambers and the volume of blood being pumped. This confirmed that the two signals are complementary; the sound tells the story of the flow, while the pulse tells the story of the volume and the vessel's resistance. The researchers emphasized that their work does not claim to have solved heart disease diagnosis. The system was tested on a specific group of patients in a single hospital, and its ability to work in different environments or with different populations has not yet been proven. Furthermore, the predictions come with a wide margin of error, meaning they cannot be used to make fine-grained clinical decisions on their own.
What this research achieves is a significant step forward in repurposing an old tool for a new job. It proves that the physical information contained in a simple heart sound and a finger pulse is richer than previously thought. By using artificial intelligence to decode these signals, the researchers have created a method that can provide a quantitative snapshot of a patient's heart health in seconds. While it is not a replacement for the detailed view provided by an ultrasound, it offers a way to bring that level of insight to the point of first contact. In a world where heart disease is often silent until it is too late, a tool that can reliably flag the most at-risk patients for immediate care represents a meaningful shift in how we approach heart health. The technology is not magic, nor is it a perfect solution, but it is a concrete demonstration that the stethoscope, when paired with modern computing, can do much more than just listen.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.