An electrocardiogram is like a recording of an engine's hum: an experienced mechanic can hear when something's off. A heart ultrasound shows the engine itself from the inside, but the machine is pricey and specialists are scarce. The new model bridges both approaches: it learns from millions of ordinary ECGs and flags those whose hearts have already changed, even if the person feels nothing.
Previously, such programs were black boxes: you got a diagnosis, but no clue why. Now they use a good old statistical framework topped with modern neural network detectors. It turned out to be 1% more accurate across all metrics, and that's with only a third of the data needed for training. Testing on different patient groups confirmed the method works reliably.
The study used 80,000 pairs of ECGs and ultrasounds: that's like a cardiologist comparing heart recordings from the population of an entire medium-sized city.
In the series "Westworld," machines read people via biometrics; here, an algorithm reads the heart via electrical signals — almost like an emotion detector, but for diseases.