AI Watch
As AI-driven remote patient monitoring scales across Medicare, a critical question remains unanswered: are these algorithms working equally well for everyone? Research published in 2026 confirms that AI models built on decades of unequal healthcare utilization data can replicate and amplify those inequities — producing less accurate risk alerts for low-income, rural, and minority patients who are the very populations remote monitoring is supposed to serve.
The bias-amplification problem
AI algorithms trained on historical healthcare data inherit the patterns of that data — including patterns shaped by decades of differential access, undertesting, and undertreatment in minority and low-income communities. A NIH PMC study published in 2026 on bias-mitigated AI in health systems concludes that standard training approaches reproduce existing disparities unless deliberately corrected with demographic subgroup validation and reweighting techniques. In RPM specifically, if a heart failure monitoring algorithm was primarily validated on white, insured, urban patients, it may generate fewer and less accurate early-warning alerts for elderly Black patients in rural Mississippi — precisely the patients at highest clinical risk.
The regulatory gap
The FDA’s dominant clearance pathway for AI medical devices — 510(k) — requires only “substantial equivalence” to a predicate device, not proof of clinical superiority or demographic equity. Of the 950+ AI/ML devices cleared through mid-2024, the vast majority used this pathway. There is currently no binding federal requirement that AI-enabled RPM software demonstrate equal performance across age, race, income, or geographic subgroups before receiving clearance.
CMS’s February 2026 Request for Information on AI tools in Medicare administration explicitly flagged equity as a required disclosure area — a notable signal. But an RFI is not a rule, and no binding bias audit standard currently applies to RPM software. H.R.7064 — the AI in Health Care Efficiency and Study Act — would mandate a federal study of AI outcomes in clinical settings by demographic subgroup, but the bill remains pending in the 119th Congress.
What accountability would look like
Meaningful accountability would require, at minimum: (1) mandatory demographic subgroup performance disclosure as a condition of FDA 510(k) clearance for AI-enabled RPM devices; (2) CMS coverage conditions requiring vendors to demonstrate equity benchmarks before Medicare reimbursement is available; and (3) post-market surveillance requirements analogous to those applied to pharmaceuticals. Until such standards exist, Medicare beneficiaries in underserved communities may be enrolled in monitoring systems that were never fully tested for people like them — and neither they nor their physicians would have any way to know.