What good looks like: five principles of rural predictive AI
State agencies are choosing their H.R.1 AI partners now. Here’s how to evaluate them.
Author Dr James Daniell, Chief AI Officer Orchestral | Reviewed by Dr Jack Vojak M.D. and Dr Luke Boyle, Honorary Senior Lecturer, University of Auckland
State agencies that received Rural Health Transformation Program awards in December face a genuinely hard job. The program sets five strategic goals and a technical scoring methodology CMS re-evaluates annually. The question they are now putting to technology predictive AI vendors is not unreasonable: can you show measurable outcomes against that methodology, year on year?
Most of the platforms they already know were not built for that question. That is not a criticism. The first wave of AI in healthcare was designed to help individuals work through existing workflows faster, and it succeeded at that job. Documentation, decision support, administrative processing: real problems, genuinely solved.
But there is an important difference between helping individuals move faster and helping institutions work differently. Early factories did not get much out of switching from steam to electricity when they kept the same floor plans. In rural health, the redesign runs on predictive analytics.
H.R.1 is asking state agencies to redesign how care reaches rural populations: who gets identified, by what signal, and what happens next. The platforms that will matter in this program are built for that brief.
Principle 1: Individual-level prediction, not cohort dashboards
Population health platforms show you the risk. Individual prediction tells you who to call.
Population health dashboards tell you about groups. For example, they may show you that 23 percent of your rural Medicaid members with diabetes have an A1C above 9%, or that readmission rates in your Critical Access Hospital (CAH) network are running four points above the national average. That information is accurate. It does not tell you who to call.
Care teams do not treat populations. They treat patients. And in rural west Texas, where a single care coordinator may cover a service area larger than most east coast states, the question that matters is not “which cohort is at risk”. It is "which patient do I call this afternoon?”
Individual-level prediction answers that question. It surfaces the specific member whose claims, lab, pharmacy, and social determinants of health data — flowing through networks like Healthconnect Texas — indicate that an intervention in the next 72 hours changes what happens to them.
That kind of precision carries a responsibility. A risk score is a care signal, not a verdict. The governance layer around any individual-level model has to make that distinction real: who sees the score, for what purpose, with what accountability at the point of action. A platform that identifies a high-risk individual to support them is a fundamentally different thing from one that identifies them to manage or penalize them. That difference has to be built in, not assumed.
In the above example, Orchestral draws those signals from across three data sources: clinical measurements including weight and height, lab results including the most recent A1C, and Medicaid claims activity. The model combines them to produce a prioritized patient list — not a risk band, but a ranked order. If a care team has capacity to reach ten patients this week, they see the ten highest-priority individuals from their population, with enough context to act. The output answers the question the section opened with: here is who to call this afternoon.
For state agencies, this is what the population health clinical infrastructure scoring factor is actually measuring. Not dashboards. Individual-level clinical impact at scale.
Principle 2: SDOH-native and rural-aware
Most predictive models were built on urban data. Applied to rural populations, they predict something else.
Most predictive models in healthcare were built on urban and suburban data. Applied to rural Medicaid members, they produce risk scores that do not reflect how rural health actually works . The factors that drive outcomes in rural populations are structurally different.
Distance to specialist is not a social inconvenience in rural health. It is a clinical variable. A prediabetic patient 80 miles from the nearest endocrinologist, with no reliable transportation and seasonal agricultural employment that makes weekday appointments impossible, carries a materially different risk profile than a patient with the same A1C in a city. A model that does not account for that difference is not appropriate in a rural health setting.
Social determinants of health data like housing stability, food security, transport access, broadband availability, employment type, need to be native to the model architecture, not appended after the fact. And the model needs to be calibrated against rural populations, not retrofitted from data collected somewhere else.
A rural-aware model goes beyond standard clinical measures. Variables like distance to the nearest specialist, whether a patient can access fresh food affordably, or local infrastructure gaps are not peripheral context. These are predictive signals. A diabetes prediction model that accounts for the reality of rural life reflects risk in a way that a model built on urban data cannot.
For state agencies, this maps to the data infrastructure scoring factor — and the rural data environment makes it a harder test than it sounds. A platform that produces no useful output when the Health Information Exchange (HIE) has incomplete claims coverage is not a rural platform.
Principle 3: Interoperable with the infrastructure rural providers already have.
A platform that needs ideal data conditions before it can work is not a rural platform.
Rural health systems work with what they have. Many are running EHR systems implemented a decade ago. Some are connected to state HIEs with partial claims coverage. Some are not connected at all. That is not a failure of ambition. It is the reality of building healthcare infrastructure in under-served communities with constrained resources.
A predictive AI platform that needs clean, complete, standardized data before it can do anything useful is not a rural platform. It is a platform built for conditions rural providers do not have.
Real interoperability means two things. First, it connects to what is already there, like legacy EHR systems, state HIE feeds, pharmacy and lab data without requiring the provider to rebuild their data environment first. Second, it keeps working when the data picture is partial. A risk model that goes quiet because claims data is missing from a rural county in the network that has just stopped serving the patients in that county.
When a state HIE runs Orchestral, the rural providers connected to that network gain access quickly, without each one requiring a separate integration project. The platform ingests any data type, which matters in rural environments where data standardization is inconsistent, and formats vary across legacy systems.
For state agencies, this is the practical test of the IT advances scoring factor. A platform that only performs under ideal data conditions is not a rural platform.
Principle 4: Agents that act, not insights that sit
An alert waiting for someone to notice it is not a care intervention.
A dashboard that flags a high-risk patient is only useful if someone sees the flag, judges it correctly, and does something about it in time. In a rural care team managing hundreds of patients across a large geography, that chain breaks more often than it holds.
The distinction is not technical. It is operational. An alert is a notification. An agent acts. When Orchestral identifies a patient at elevated risk, it does not wait for a clinician to notice a flag, it books the lab test, schedules the pharmacy consult, or routes the intervention to the right care team member with enough context to act. It pulls data from multiple sources, runs the calculation, and moves. Out of dashboards and into workflows.
That capacity to act carries the same responsibility as individual-level prediction. The action logic and the human accountability at each step have to be defined before the agent runs, not after something goes wrong.
For state agencies, this is what helps them demonstrate progress on innovative care goal: coordinated care, measurable outcomes, and care shifted away from the most expensive settings.
Principle 5: Governed and measurable, mapped to the CMS scoring factors
Technical scoring factors are directly addressable by predictive AI. Here is how they map.
The principles above describe what a rural predictive AI platform should do. This one describes how you prove it.
CMS allocates the programme’s $25 billion workload funding using rural and technical score factors, with workload funding allocations recalculated annually based on state reporting and programme performance1. Five program areas are directly addressable by predictive AI, and they map to the principles above. What follows is Orchestral’s mapping of the Notice of Funding Opportunity’s (NOFO) technical factor list, not language CMS uses verbatim; the primary source is the NOFO itself. But the mapping is worth making explicit, because it turns a capability conversation into an accountability one.
The factors, and why they matter in this program:
Population health clinical infrastructure: The broadest factor. It asks whether your platform can demonstrate individual-level clinical impact across your rural Medicaid population. Principle 1 addresses this directly.
Data infrastructure: Rural data environments break platforms built on urban assumptions. This factor rewards FHIR-native, HIE-connected platforms that keep working when the data picture is incomplete.
Consumer-facing technology: RHT explicitly funds it. In rural health, when a specialist is 80 miles away, the patient-facing channel is not supplementary. It is primary.
Remote care services: The delivery mechanism for diabetes prevention, post-discharge monitoring, and maternal health. Without it, rural reach collapses to whoever can get to a clinic.
Innovative care models: The scoring-factor expression of Principle 4. A platform that produces reports scores here. A platform whose agents change care pathways scores higher.
What this looks like in practice
Lisa, 47, rural west Texas Prediabetic, no local diabetes service, 70 miles from the nearest endocrinologist.
Lisa works in seasonal agricultural work in rural west Texas and was flagged as prediabetic at her last primary care visit. Her primary care provider is 40 minutes away. The nearest endocrinologist is 70 miles in the other direction. There is no in-person community diabetes service within practical reach, and her work schedule makes weekday appointments difficult.
Without structured prevention support, 5-10 percent (sometimes up to 11 percent) of people with prediabetes progress to Type 2 diabetes each year2. For Lisa, that trajectory means higher long-term costs for the state, worse outcomes for her, and a care system that had the information to act earlier but didn’t.
The Orchestral diabetes and prediabetes prevention agent identifies Lisa before that progression advances. Her risk stratification draws on claims data, lab results, pharmacy history, and social determinants of health signals, including the distance-to-specialist and employment pattern data that an urban-trained model would miss. She surfaces as high priority.
What happens next is not a dashboard alert waiting for someone to notice it. A diabetes prevention service aligned digital program is delivered to her at home. The content reaches her in the channel she uses. Her primary care provider receives a care alert with enough context to act. Her progress — engagement, A1C trajectory, course completion — is tracked back to the state plan.
The evidence for intervening early is clear. Completing a structured diabetes prevention program reduces the risk of progressing to Type 2 diabetes by 58 percent. For adults over 60, that figure rises to 71 percent3. Her progress is tracked back to the state plan: engagement, A1C trajectory, program completion.
The same platform addresses a different risk further along the disease journey. Rural patients managing Type 2 diabetes are disproportionately likely to have polypharmacy, reduced renal function, and no specialist input into their prescribing. A prescription review surfaces the error patterns that generate avoidable admissions, without requiring a specialist pharmacist at every point of prescribing. Modeling that intervention across New Mexico’s Type 2 diabetic prescribing population estimates 105 hospital admissions prevented and $2.3 million in direct hospital costs avoided in a single year. (Based on Orchestral’s internal modeling of rural health data, June 2026)
For a state agency, this is what measurable outcomes against the CMS scoring methodology actually looks like.
Ten questions to take into your next vendor conversation
Each one maps to one of the principles above and to what CMS is actually scoring.
The principles above are a framework for evaluation, not a checklist. Certainly if you are in active vendor conversations, these are the questions worth asking.
Can you show us an individual-level risk prediction — not a cohort view — and walk us through what signals generated it?
What SDOH data sources do you ingest, and how is your model calibrated specifically for rural populations?
What happens when the HIE feed is incomplete or claims data is missing?
What does a rural provider need to have in place before you can go live, and how long does integration take?
When your platform identifies a high-risk patient, what specifically happens next — what does the agent do, and what does the care team member see?
How do you ensure a risk score is used to support a patient, not to penalize or manage them?
What governance controls are in place and who sees the score, for what purpose, with what audit trail?
How do you demonstrate measurable progress against the AI-relevant CMS technical scoring factors?
What outcome metrics do you track, and how do they map to the five H.R.1 statutory goals?
What does a pilot structure look like in terms of timeline, data requirements, go-live conditions?
The platforms that earn state agency trust will be the ones that can answer these questions clearly, with evidence, before the contract is signed.
Citations
Centers for Medicare & Medicaid Services. (2025). Rural Health Transformation Program (CMS-RHT-26-001): Notice of Funding Opportunity. U.S. Department of Health and Human Services. https://simpler.grants.gov/opportunity/782f996f-78f8-4742-8b68-d2bf50c87f99.
Alizadeh Z, Baradaran HR, Kohansal K, et al. (2022). Are the determinants of the progression to type 2 diabetes and regression to normoglycemia in the populations with pre-diabetes the same? Frontiers in Endocrinology, 13, 1041808. https://doi.org/10.3389/fendo.2022.1041808
Knowler WC et al. Reduction in the Incidence of Type 2 Diabetes with Lifestyle Intervention or Metformin. New England Journal of Medicine. 2002;346(6):393-403.