Generative AI is not replacing your QA team. It is making the good ones faster, sharper, and harder to fool. That is the honest takeaway from a recent conversation with Richard "RJ" Kedziora, our very own co-founder and CEO of Estenda Solutions, who sat down on the Test Guild Automation Podcast to talk through how generative AI in healthcare is changing software testing quality.

If you build or test healthcare software, you already know the stakes are not the same as everywhere else. A bug in a shopping cart is annoying. A bug in a system reading someone's blood glucose or heart rate can cause real harm. RJ puts it plainly: "Testing and quality are at the heart of everything we do, for the reason you just said. It is about people, it's about their health and wellness." That is the lens for everything you are about to read.

How Generative AI Is Improving Healthcare Software Testing Quality with Richard Kedziora

Treat Generative AI as an "80% Solution," Not an Oracle

Treat every AI output as a strong first draft that still needs a human check. That is the single most important habit RJ recommends. The tool is fast, and it sounds confident, which is exactly why it is dangerous when you stop paying attention.

RJ frames it in a way that sticks. "I think of it as an intern," he says. "Think of it as that first-year person. It's very capable, but it also sounds very trustworthy, sounds very authoritative. When it just spits out something, it's very easy to go, OK, that's right. Now you have to go back and double-check it. We do use it a lot. I'm a huge advocate for it, but it's an 80% solution."

The remaining 20% is where your expertise lives. He points to the well-known cases where professionals trusted AI blindly and paid for it. "Never just sit here and be like, oh yeah, OK. There are various stories where lawyers have gotten in trouble because they just took it at face value and didn't cross-check the information." The lesson for QA is simple. Use AI to move faster, then verify as the outcome matters, because in healthcare it does.

Test for Hallucinations, Bias, and Non-Determinism

When the system under test is itself a generative model, you have to test for three specific failure modes: hallucinations, bias, and non-determinism. Most traditional test plans miss all three, and healthcare AI testing lives or dies on catching them.

RJ lays them out clearly. First, hallucinations. "Is it just making stuff up? So you've got to be aware of that as you're testing, and how do you mitigate against that?" Second, bias in the training data. "If you start out with a generic large language model and then you fine-tune it and train it, add in your specific information, is it accurately representing that information? You've got to be careful of that as well."

Third, and this one reshapes your whole approach: non-determinism. "The challenge here is it's non-deterministic. I can ask the same question and five minutes later ask the same question again and get a different answer. So it's not just testing; it's also monitoring over time to make sure it is performing as you expect." Read that last line twice. Quality assurance for AI is not a one-time gate. It is continuous monitoring, which lines up closely with frameworks like the NIST AI Risk Management Framework.

Validate Algorithms Against Representative Populations

An algorithm that works in one hospital can quietly fail in another. So you have to validate it against data that actually represents the population it will serve. This is a quality requirement, not a nice extra, and the FDA has flagged it too.

RJ is clear here. "You can't just take an algorithm that was generated in one hospital and move it into another hospital without testing it. What are the specifics of your population?" He uses geography to make it real. "I live in the suburbs of Philadelphia, and there's a very specific demographic of people that live here. Now I take that and move it down South, and there's a different demographic. Or move it out West, or Canada, or Mexico, or Japan. Do the same algorithms that were developed here work for those populations? Probably not."

The fix is to build with breadth from the start. "That's why you need to test, and as you're developing, make sure you're using a data set that is representative of a wider population." If you are building software as a medical device, this is also central to how the FDA evaluates AI enabled tools. Estenda's healthcare and software research team designs validation studies with exactly this in mind.

Use AI to Generate Test Cases from Screenshots and Requirements

You can hand a vision-capable model a screenshot of your interface, and it will write test cases for it, including edge cases and security checks a newer tester might miss. This is one of the fastest wins in AI test case generation, and it genuinely surprised RJ the first time he tried it.

"I've done this a couple of times now," he says. "You take the screenshot, drop it into one of the models that has vision capability, and say, ' Write test cases for this. That's where it truly amazed me months ago when I first did this." The model reads the design like a person would. "Typically, if you have a data input field and you put a little asterisk next to it, which means it's a required field, it knew that. It was writing test cases that the username is a required field."

It goes further than surface checks. "It even helps with security testing, like SQL injection. If you're newer to the field, maybe you don't remember to test SQL injection, or you're not familiar with it. So you can just ask, what is SQL injection, how does it work, to really help you out." That is training and coverage in one step. Not bad for a screenshot.

Generate Realistic Synthetic Test Data (Especially for Wearables and EMR)

Generative AI is excellent at creating realistic synthetic test data, which is a huge deal in healthcare where you need volume without touching real patient records. Wearables and electronic medical record systems produce enormous, varied data, and testing them properly means feeding them equally varied inputs.

"Probably last is just test data generation," RJ says. "Particularly in healthcare, with the wearables, with the electronic medical record systems, there's so much data that you need to test these systems. It's beautiful for creating data for you."

His heart rate example shows how nuanced the output can be. "I said, generate me a heart rate profile of a person running a mile, and it generates that data. For every minute, it starts out at, say, 90 beats per minute, goes up, and then toward the end of the run it comes back down. Then I said, generate the same data for a 30-year-old athlete in really good shape, and it drops the heart rate down." Here is the part that got him. "It doesn't understand heart rate. It doesn't understand exercise parameters or age. But it was able to generate a good representative data set that differentiated between an out-of-shape 60-year-old and an in-shape athlete. So it was very meaningful to help test systems." Estenda builds these kinds of pipelines through its healthcare data analytics services.

Keep a Human Review Process and Ask "What Did I Forget?"

The highest value question you can ask an AI during testing is "what did I forget?" It turns the tool into a second set of eyes on your own blind spots, without ever removing the human review that regulated software demands.

RJ ties this straight back to Estenda's documented process. "I always love asking the question, what did I forget? What is that other thing that I forgot? Our process is very SOP driven, template driven, checklist driven, to make sure we're not forgetting the little things you need to think about when developing a system. But use the AI to help you with that."

He even points the model at recognized standards to pressure test his own thinking. "There's a standard, ISO 25000, where it talks about scalability and usability and things like that. As I'm writing those prompts, I say, based on this ISO standard, which covers all these things, what am I not remembering to ask?" That is a smart move. You are combining a quality framework like ISO/IEC 25010 software quality characteristics with AI's recall, then keeping a human in the loop to decide what matters.

Use AI Earlier in the Lifecycle to Build Better Requirements

Quality does not start at testing. It starts at requirements, and RJ uses AI right there at the beginning to build better inputs faster. Better requirements mean better tests, so this is testing quality even before a single test case exists.

"I've been doing user interviews for decades," he says. "By entering my parameters into ChatGPT or Claude, it's spitting them out a lot faster for me and getting me going. It's at 80%. Then I can go and tweak those questions and make them different. But it's an efficiency thing." He gives a concrete case. "I'm developing a system addressing prostate cancer at a specific institution. There's a doctor with 30 years of experience who has written these articles. You can tell it, here are these journal articles, what questions do I need to ask him to get to the core of the system?"

Then the human review kicks back in. "Once we have requirements created, we do a lot of human review, but you can also plug them into the system." AI drafts, you refine, quality compounds. Estenda's strategy consulting is built around this kind of early planning.

Design Data and Quality Strategy First, Because "Garbage In, Garbage Out"

Before you write a single prompt, you need a data strategy and a quality strategy. Skip that step and no amount of clever AI testing will save you. This is the foundation everything else sits on.

RJ says it is often the first thing missing on startup projects. "One of the first early questions is always around, what's your data strategy? What's your quality strategy? Some companies, particularly on the startup side, they don't have that figured out yet. You're not going to get very far if you don't have your data strategy and your quality strategy in place. As you start thinking about AI, it starts with the data. Garbage in, garbage out."

That phrase is old for a reason. The quality of your test outputs will never beat the quality of your inputs. If your training data is biased, incomplete, or mislabeled, your model will fail in ways that are hard to catch and easy to ship. Getting the data foundation right is exactly the kind of work Estenda's AI and ML development services are designed to support.

Understand HIPAA Risk Without Overreacting to It

HIPAA risk with generative AI is real, but RJ considers the fear of leaked patient data overblown when you manage identifiers and contracts properly. The goal is controlled usage, not avoidance.

"In my opinion, it's one of the overblown concerns," he says of the idea that a model will spit out a real patient's data. "I'm not saying it's never going to happen, but it's an overblown concern." His practical guidance is to strip identifiers. "You don't have to put an identifier in there to get meaningful feedback from it." A physician can ask a clinical question without attaching a name.

The bigger control is contractual. "These corporations are now signing business associate agreements. They weren't early on, but they are now, and they hold your data separately, so it's not incorporated into the model." His caution is to check the fine print each time. "As they're rolling out new models, make sure you know what your business associate agreement actually covers. Does it cover the model that just came out last week? Probably not." Smart HIPAA compliance is about reading the agreement and managing risk, which you can learn more about directly from HHS on HIPAA.

Ready to Build Testing Quality Into Your Healthcare Software?

You do not have to figure out AI testing, data strategy, and regulatory validation on your own. At Estenda, we work with MedTech, life sciences, and digital health organizations to build software that holds up to real-world use and regulatory scrutiny. Whether you need custom software development, AI validation, or a clear quality and data strategy, our ISO 13485 certified team helps you move fast without cutting corners.

Book your free 30-minute consultation today, or contact Estenda at info@estenda.com.

Want more of RJ's thinking on responsible AI? Read our related article on what responsible AI looks like in clinical research.

Frequently Asked Questions

How does generative AI improve healthcare software testing quality?

Generative AI improves testing quality by generating test cases from screenshots and requirements, creating realistic synthetic test data at scale, and flagging gaps a tester might forget. It speeds up repetitive QA work while humans keep responsibility for verification, which matters most in regulated healthcare software.

Can you trust AI-generated test results in healthcare?

Not without human review. RJ Kedziora describes generative AI as an "80% solution" that sounds authoritative but still needs cross-checking. In healthcare, where errors can affect patients, a person must validate every AI output before it is trusted.

What should you test for when the software itself uses AI?

Test for three things: hallucinations, where the model invents information; bias in the training data; and non-determinism, where the same input can produce different answers. Because outputs can drift, AI testing also requires continuous monitoring over time, not a single test pass.

Is it safe to use generative AI with patient data under HIPAA?

It can be, with the right controls. Remove personal identifiers before prompting, and confirm that the AI vendor has signed a business associate agreement covering the specific model you are using. Many large language model providers now sign these agreements and keep your data separate from training.

Why shouldn’t you blindly trust  a healthcare algorithm across different populations?

Because patient populations differ. An algorithm trained on one demographic may perform poorly on another, so you must retest it against data that represents the new population. Validating against representative data sets is essential for both quality and regulatory approval.

Does generative AI replace QA testers in healthcare?

No. It changes their role. As RJ puts it, "It's not going to replace you, but people using it are going to replace you. So get to know it." AI handles scale and speed, while testers own judgment, verification, and accountability.