Detection of ecg anomalies(especially episodic ones with intermittent recording) was the subject of the physionet cardiac computing challenge almost 10years ago[0].
It’s amazing how far machine learning has come. I know teach a version of this challenge as a one day in class activity in my department’s physiology class. They actually get to train multiple models on a gpu cluster (and compare that to trying to train models on their laptops).
One thing we reinforce in the lesson is human vs. computer “interpretation”. They/clinicians can look at ecgs and make some sense of them. An LTSM is worse than random chance/a medical student. However moving to the frequency domain makes the LTSM more accurate than cardiologists, but neither they nor clinicians can “see” afib ina spectrograph. It’s a great way to talk about algorithmic versus human reasoning and illustrate that to students.
That then gets reinforced with other case studies of the ying and yang of human and machine decision making throughout our curriculum- like alpha fold working great until you ask it about a structure in the absence of oxygen, because that’s not in its training data.