AI Meets Biosensing
The hard problem in health AI is not the model. It is everything that happens before the model sees the signal.
Health AI papers tend to begin after the interesting part is over: a clean dataset, artefact removed, labels agreed. In a device that is worn continuously, none of that is given.
Signal before intelligence
Motion artefact, electrode contact, ambient interference and the simple fact that people live their lives while wearing the sensor — these dominate the error budget. A model trained on clean data and deployed on lived data will fail in ways that look like confidence.
So the sequence matters: signal → sensing → AI → interpretation. Each arrow is a place where quality is either preserved or lost, and no later stage can recover what an earlier one destroyed.
Designing hardware and models together
When the sensor and the model are designed by the same team, the model can rely on properties the hardware guarantees, and the hardware can drop what the model does not need. That co-design is the advantage — and it is unavailable to anyone who only has one half.