An electrocardiogram is like a long strip of paper on which the heart leaves its signature. A doctor studies it to catch rhythm disturbances, but when the recording lasts for days, the strip stretches for miles. The eye tires, attention wanders, and rare but dangerous irregularities can slip by.
AI-HEART works like a tireless assistant that scans the entire strip without a break. It takes multi-day recordings from three leads (like three cameras capturing the heart from different angles), cleans out noise, finds the boundaries of the waves, and recognizes different types of arrhythmias. To teach it to distinguish rare rhythms, the developers not only used recordings labeled by doctors but also created additional examples using generative models, as if you were training facial recognition by filling in missing portraits.
The doctor is not left out: they review the system's conclusions, correct errors, and these corrections feed back into training. This way, the platform gradually learns from its own mistakes and becomes more accurate. The main result is not just high accuracy but the ability to integrate such a service into daily clinical practice: store recordings, track every decision, and conduct audits with confidence.
Such a system is not a replacement for a cardiologist but a third eye that never blinks. A strip spanning several days is no longer overwhelming, and a rare rhythm glitch gets a chance to be noticed in time.
To train the neural network on rare arrhythmias, generative models were used to create synthetic examples, like an artist filling in missing frames in a film.
In the series 'Westworld,' machines constantly analyze people's biometrics to predict behavior. AI-HEART takes the first step toward such continuous monitoring, but for now only for the heart.