AI-Enabled Digital Stethoscope for Early Detection of Congenital Heart Disease: Prospective Study

Every year, thousands of children are born with structural defects in their hearts. In many parts of the world, finding these problems early is a matter of life and death, yet access to specialized doctors and expensive imaging machines is often scarce. For generations, the first step in spotting these issues has been the simple act of listening. A doctor places a stethoscope against a child's chest, listening for the rhythm of the heartbeat and the presence of any extra sounds, known as murmurs, that might signal a problem. This method is free, non-invasive, and available almost everywhere, but it relies entirely on the skill and experience of the person holding the device. A murmur that is obvious to a specialist might be missed by a general practitioner, and the accuracy of the diagnosis can vary wildly depending on who is listening.

To bridge this gap, researchers have begun exploring how computers can learn to listen better than humans. By combining a digital stethoscope with artificial intelligence, scientists hope to create a tool that can hear the subtle signs of heart disease with the same consistency as a top expert, regardless of the environment or the listener's training. This approach does not replace the need for advanced imaging, which remains the gold standard for confirming a diagnosis, but it offers a powerful way to decide who needs that advanced care immediately. A new study from a hospital in India puts this idea to the test, asking whether a computer program can reliably detect congenital heart disease in children just by analyzing the sounds of their hearts.

The researchers set out to build and test a system that could listen to heartbeats and decide if a child has a heart defect. They recruited 721 children and young adults, ranging from infants to adults, visiting a specialized cardiac unit. About half of these participants had confirmed heart defects, while the others had structurally normal hearts. Using a digital stethoscope that records sound and sends it wirelessly to a smartphone, the team captured heart sounds from four standard spots on the chest: the upper right, upper left, lower left, and the tip of the heart. Each recording lasted twenty seconds. The team then split their data, using sounds from 301 participants to teach a computer program what a healthy heart sounds like versus what a defective one sounds like. The remaining 420 participants were used to test the program without the researchers knowing the true diagnosis beforehand, ensuring the test was fair and unbiased.

The computer program worked by turning the raw sound waves into a visual map called a mel-spectrogram. This map displays sound in a way that mimics how the human ear perceives pitch, allowing the computer to see patterns in the noise that are invisible to the naked ear. The program analyzed these maps using a type of artificial intelligence known as a convolutional neural network, which is designed to recognize complex patterns in images and sounds. The system learned to identify specific acoustic signatures associated with different types of heart defects, such as the turbulent flow of blood through a narrowed valve or the continuous rush of blood through a hole in the heart wall. When the program heard a sound, it calculated a probability score to determine if the heart was likely abnormal.

When the researchers tested the system on the 420 unseen participants, the results were striking. The artificial intelligence correctly identified 222 out of 233 children who had heart defects, missing only 11 cases. It also correctly identified 171 out of 187 children who had healthy hearts, flagging 16 healthy children as potentially having a problem. This performance translated to a sensitivity of 95.2 percent and a specificity of 91.4 percent when compared against the final diagnosis made by echocardiography, the ultrasound imaging used as the definitive reference. In simpler terms, the computer missed very few sick children and rarely raised a false alarm for healthy ones. The system performed just as well as a trained pediatric cardiologist listening with a traditional stethoscope, and in some cases, it even outperformed human listeners in noisy environments or with uncooperative children who were crying or moving.

The study also looked closely at the types of heart defects the system could find. It was exceptionally good at spotting defects where blood flow was increased, such as holes between the heart chambers, catching 98 percent of these cases. It was also highly effective at finding defects where blood flow was decreased, such as narrowed valves, with a 93 percent detection rate. Even for defects where blood flow remained normal, the system maintained a strong detection rate of 91 percent. The researchers noted that the few cases the system missed were often those where the heart defect produced very faint or no audible sounds at all, even for an expert human doctor. This suggests that the limitations of the computer were similar to the limitations of human hearing, rather than a failure of the technology itself.

The performance of the system remained consistent across all age groups, from infants under one year old to adults over eighteen. This is particularly significant because listening to the hearts of infants is notoriously difficult; their hearts beat much faster, and they are often crying or restless, making it hard for a human to isolate the sound of a murmur. The fact that the computer maintained high accuracy in these challenging conditions suggests it could be a vital tool for newborn screening programs. The researchers acknowledged that their study was conducted at a single center and that the system was trained on data from children who had not yet undergone heart surgery. They also noted that the system was designed to detect the presence of a problem, not to identify the specific type of defect or its severity.

Despite these limitations, the findings offer a compelling vision for the future of heart disease screening. The study demonstrates that an artificial intelligence system integrated into a digital stethoscope can detect congenital heart disease in children with high accuracy, matching the performance of expert specialists. By providing a reliable, non-invasive way to screen for heart defects, this technology could help identify children who need urgent care much earlier, especially in regions where access to specialized cardiologists is limited. The researchers conclude that such a tool has the potential to standardize screening, reduce delays in diagnosis, and ultimately improve outcomes for children with heart defects around the world.