Why a General AI Beats Task-Specific Models at Predicting Vascular Age

JUST IN: A task-specific AI model designed to predict vascular age is underperforming. It's trained on 212,231 UK Biobank subjects, yet flops when applied to a different clinical group. The predictions land in a limited 38-67 year range, regardless of the actual age. Meanwhile, a general-purpose AI, with no special training for age, nails the task with lower error rates. How's that for a twist?

The Unexpected Champion

Three open-source models were put to the test on 906 surgical patients from PulseDB. The generalist, Pulse-PPG, scored a mean absolute error (MAE) of 9.28 years. It outshone the specialized AI-PPG Age (9.72 years in Probe Mode) and even models combining heart rate/heart rate variability with demographics (9.59 years). The kicker? Throw in some demographic data, and Pulse-PPG drops its MAE to 8.22 years.

Data Size: The Elephant in the Room

So what's causing the gap? Apple's PpgAge model boasts a MAE of just 2.43 years. The stark difference isn't due to superior architecture but rather the sheer size of the dataset used (213,593 subjects versus our 906). Plus, the population differences play a big role. Our learning curve indicates there's still room to improve with more data.

Why You Should Care

The labs are scrambling to rethink how we approach biological age prediction. If a jack-of-all-trades AI can beat a specialized model, what does that say about our current approach? Maybe it's time to question whether hyper-specialized models are always the way forward.

And just like that, the leaderboard shifts. The real question is: will bigger datasets eclipse the need for tailored training? It's a wild thought, but it's one the industry can't ignore.

Sources confirm: the code is publicly available for anyone curious to dive deeper. With results like this, the medical AI field is far from predictable. Who knows what the next benchmark will reveal?