Diabetic retinopathy is the leading cause of blindness in working-age adults, and about 90 percent of that vision loss is preventable when it is caught early. Catching it means reading retinal images, and there are not enough specialists to read them all.

At Estenda, our Co-founder and COO, RJ Kedziora, guested on the AI For Pharma Growth podcast to talk about how AI can close gaps like this for patients the system keeps missing. RJ has spent more than two decades building digital health tools, and his view is clear. "Technology is important," he says, "but it is as much about people and process as opposed to the technology itself." These are five things it takes to build AI that reaches underserved patients, in his words.

What Does It Take to Build AI That Includes Underserved Patients?

Scale Specialist Judgment Where Specialists Are Scarce

"One of our bigger projects is with the Indian Health Service, which is responsible for the Native population here in the United States," RJ says. Diabetes runs high in that population, and diabetic retinopathy is a leading cause of preventable blindness. "So if you find it, you can make an impact."

Finding it means getting a trained eye on a retinal image, and for remote and underserved patients, there is often no specialist anywhere nearby. The IHS teleophthalmology program we supported closes that distance by having specialists read the images remotely. The program imaged more than 120,000 patients between 2000 and 2021, and over that period, retinopathy rates fell to levels comparable to other populations. Every image still needs a human to read it, though, which limits how far one program can stretch. AI extends that reach further. It reads the images in a first pass and flags the ones a clinician needs to see, so the same specialists can cover far more people than they could one image at a time. 

Confront the Bias Sitting in Your Training Data

A model learns from the data you give it, and healthcare data carries the same gaps the system always has. "There was just a recent research report that women are still underrepresented in clinical trials," RJ notes. As of 2019, women made up only about 40 percent of participants in trials for cancer, cardiovascular disease, and psychiatric disorders, three conditions that affect them heavily.

FDA reviews for AI-enabled devices do call for representative data and testing across patient subgroups, yet demographic reporting on cleared devices stays thin, and a model that performs well at one site can still underperform at another. Building for any population means testing for bias across the populations you intend to serve, rather than assuming a model that worked once will work everywhere.

Solve the Data Access Problem Before You Touch the Model

The hardest part of healthcare AI is rarely the algorithm. "One of the biggest challenges in AI today is access to that data, understanding that data," RJ says. Two things get in the way.

The first is a misunderstanding of HIPAA. People treat it as a reason not to share data, RJ says, but that was never its purpose. With patient permission and proper privacy safeguards, that data can be used.

The second is actually reaching underrepresented communities. They are the least visible in the data and often the ones a well-built tool can help the most, and including them takes trust and outreach, and it is not just a legal sign-off. "Reaching those populations, here in the US and around the world, is very much an educational and marketing challenge," RJ says. That makes data access a people and process problem as much as a technical one.

Meet Patients on the Devices They Already Carry

Most patients are not in a clinic. They are holding a phone. "A lot of those engagements with the patient are on the phone," RJ says. "We are all carrying around our smartphones every day." That device is how care reaches people who live far from a provider or cannot take time off to see one.

Short visits miss most of a patient's life. "You go to the doctor, you are there for seven to ten minutes, 15 if you are lucky," he says. "So much of what you do outside of the four walls makes a difference in your life." Combining wearable information with remote monitoring closes that gap, giving practitioners a view of what happens between visits and giving AI the data to act on. 

Digital therapeutics push this further, giving people validated, sometimes prescribed software to manage mental health and chronic conditions from the device in their pocket. As RJ puts it, "a couple dozen of those can make a real difference in helping people in the real world."

Speak the Patient's Language, Not the Algorithm's

Data only helps a patient who can understand it. Language level is where this succeeds or fails. "You really need to be cautious about the language level that you are writing at," he says. "And then, as you translate materials, ensure that it is translated appropriately. The AI systems out there can do this very well. Still, always double-check, cross-check." A tool built this way can review its own communication for reading level and bias and adjust, which is a quieter form of empathy than most people expect from software. It is what separates a tool that technically works from one that actually reaches a patient.

Build It With a Partner Who Has Done It

Every problem this article raises is one we solve for clients. We pressure-test the idea before it burns budget, build and validate AI against real patient populations, turn messy health data into something usable, and fit the result into a workflow clinicians will actually use. We have done this with MedTech, life sciences, and digital health teams for more than 20 years, moving fast while keeping risk low and the patient at the center.

If you are building something meant to reach the patients others miss, we can help you get it right. Start with a free 30-minute consultation, no obligation. Get in touch, email info@estenda.com, or visit estenda.com.

Frequently Asked Questions

What does it take to build AI that reaches underserved patients?

It takes more than a strong model. You need to scale specialist judgment where specialists are scarce, confront bias in your training data, solve the data access problem before building, meet patients on the phones and wearables they already use, and communicate in language patients actually understand. People and process matter as much as the technology.

How does AI help patients in remote or rural areas?

AI extends expertise to places that lack it. Image analysis can screen retinal scans, skin images, and other data at scale, flagging the cases a clinician needs to review. Paired with smartphones, wearables, and remote monitoring, it reaches people who live far from a provider or cannot easily travel for care.

Why is bias in healthcare AI a risk for underserved patients?

A model trained on data from one hospital can carry that population's bias and fail for a different community. Because groups like women and minority populations are often underrepresented in the data, biased tools tend to work worst for the patients who already face the largest gaps. The fix is deliberate testing across the populations you intend to serve.

Does HIPAA prevent using patient data to build AI?

No. HIPAA was not designed to silo data. With patient permission and proper privacy safeguards, that data can be used to build tools that improve care. Treating HIPAA as an automatic wall keeps useful information locked away and slows down work that could reach underserved patients.

What are digital therapeutics?

Digital therapeutics are software products validated through clinical trials and, in many cases, prescribed like a medication to prevent or manage a condition, then reimbursed by insurance rather than paid for by the patient. They deliver care through everyday devices, which makes them a practical way to reach patients who cannot access traditional in-person services.