Continuous monitoring genuinely works. The catch is who's wearing the monitor — and whether the clinical-AI models learning from that data will ever see the people who need it most.
The monitor is not the problem. The monitor works.
Across 106,261 patients, remote telemonitoring cut systolic blood pressure by roughly 5 mmHg — the kind of shift that, at scale, prevents strokes.
So continuous, longitudinal data is real medicine now — and it's also the raw material teaching the next generation of clinical-AI risk models. Which means one quiet question decides who those models will work for: who is actually wearing the device?
3 predictions · ~3 minutes · no medical knowledge required
Prediction 1 of 3 · Who owns one
Wearable ownership isn't random. Which group owns them most?
Pick the group with the highest odds of owning a wearable
Uninsured adults
Rural households
Households earning over $200K/year
Ownership skews toward the wealthy, urban, educated, and already healthy. Relative to a reference group, the odds look like this — a reference line at 1.0× means "average":
>$200K income
2.0×
Rural households
0.65×
Uninsured
0.41×
Relative odds of wearable ownership, as reported in the surgeon-founder essay "The Quantified Few" (Techy Surgeon). Reference = 1.0×.
Prediction 2 of 3 · Who needs one
Now flip it. In US counties, does more uninsured track with more chronic disease — or less?
If monitoring helps people with chronic conditions, the people who'd benefit most are the ones already carrying the disease burden. So across all 3,078 US counties, how does the uninsured rate relate to diabetes prevalence?
Pick the relationship you'd expect
More uninsured → less diabetes (healthier populations)
No real relationship — they're unrelated
More uninsured → more diabetes (they move together)
CDC PLACES · 3,078 counties · 2024 release
County uninsured rate and diabetes prevalence move together: correlation r = 0.43. High blood pressure tracks the same direction (r = 0.13). The places with the least coverage carry the most disease.
So the device sits on the wrong wrist. The people generating the longitudinal data are the worried well; the people with the disease the models are meant to catch are the least likely to be wearing anything. The training set and the target population are drifting apart.
The gap · two real states
100 adults in each state. Blue = living with diabetes.
Same country, same devices for sale. But the disease burden — and the coverage that would pay for monitoring — are not evenly spread. (County-averaged CDC PLACES estimates.)
Massachusetts
3.0% uninsured · 10.1% diabetes
Mississippi
10.9% uninsured · 16.8% diabetes
Mississippi carries 66% more diabetes and nearly 4× the uninsured rate of Massachusetts. Guess which population an Apple-Watch-trained risk model has seen more of.
The 80/20
"If clinical AI learns from the people who need it least, it will work best for the people who need it least."
Every eval-harness conversation is about accuracy on a test set. But if the test set itself is the Apple-Watch demographic, a "validated" model can be biased and still pass. The missing step isn't more accuracy — it's a representativeness check: before you trust a score, ask whose data it was measured on.
The builder's move
Add one line to your eval report: the income, insurance, and rural/urban makeup of your training and test data, next to the accuracy number. It's the cheapest bias check you can ship — and right now almost nobody ships it.
CMS's ACCESS Model pays only when a share of your panel hits target — a scoring rule that quietly rewards enrolling the patients most likely to succeed. The same skew, one layer up.