AI-Driven Health: How Machine Learning Is Personalizing Longevity Medicine

By Dr. Farhan Abdullah, DO | Medical Director, Magnolia Functional Wellness | Southlake, TX

A patient sat in my office last month with her phone out, showing me an ad for a four hundred dollar test that promised to reveal her "true biological age." The photo next to it showed a radiant woman in her sixties holding a yoga pose on a beach. She wanted to know whether she should buy it. My answer took about twenty minutes, because the honest version isn't a yes or a no.

Artificial intelligence has arrived in longevity medicine, and it arrived the way most things arrive in this field, with the marketing running several laps ahead of the evidence. Some of what's being sold has real peer-reviewed data behind it. Some of it is a random number generator with a good logo.

I'm an internal medicine physician. I still work shifts at a hospital in Dallas while running Magnolia Functional Wellness here in Southlake, and that split gives me an odd vantage point on this technology. In the hospital I see algorithms that had to clear institutional review, validation cohorts, and regulatory scrutiny before anyone let them near a patient. In the direct-to-consumer wellness space, I see algorithms that had to clear a web developer. Same underlying math. Wildly different quality control.

So let's go through what machine learning is actually doing in this field right now, what the published literature supports, and where I think the line falls.

What "AI in medicine" Means Once You Strip the Marketing Off

Most medical AI worth taking seriously does one narrow thing. It finds patterns in a type of data that humans read poorly. It doesn't reason. It doesn't understand your life or your goals. It recognizes that a certain configuration of pixels or waveform features tends to travel with a certain outcome, because it was shown several hundred thousand examples until the association stuck.

That sounds unglamorous, and it is. It's also genuinely powerful in specific places, because there's information sitting in ordinary clinical data that the human eye can't pull out. A cardiologist reading your ECG is looking for a set of established abnormalities that medicine has named over the past century. A trained network looking at the same tracing is reading thousands of subtle relationships between features that nobody ever taught it to name.

The distinction I keep returning to with patients is between prediction and action. A tool that predicts something is interesting. A tool that changes what a physician does, and through that changes an outcome, is medicine. Most of what gets marketed to consumers stops at prediction, and frequently at prediction that was never validated in anyone resembling the person buying it.

So ask a blunt question of any AI health product: what did a clinician do differently because of this result, and did that change help anyone? If nobody can answer that, you're buying a number, not care.

Your Body Has More Than One Age, and Software Can Now Estimate Several of Them

The idea underneath biological age testing is old and basically sound. Two sixty-year-olds can be in radically different physiological shape, and the calendar is a crude proxy for how much wear a body has actually taken on. What's new is being able to estimate that wear from data you'd generate anyway during a routine visit.

Consider the retina. It's the only place in the body where you can look directly at blood vessels and nerve tissue without cutting anything open. In a paper published in the British Journal of Ophthalmology, Zhu and colleagues trained a deep learning model on more than 80,000 fundus photographs from nearly 47,000 UK Biobank participants, teaching it to guess a person's age from the image alone. It landed within about 3.5 years on average. Then they went looking at the people it guessed wrong. Each additional year of what they called the retinal age gap, meaning the amount by which the model overestimated someone, was associated with roughly a 2% higher risk of dying during follow-up. The retinal age gap study is worth reading if you like this kind of thing.

The same logic applies to the ECG, and here the work is even more striking. Attia and colleagues at Mayo Clinic trained a convolutional neural network on 12-lead tracings from just under 500,000 patients, then tested it on another 275,000. It estimated age with an average error of about seven years and identified sex with roughly 90% accuracy, all from a tracing that carries no explicit information about either. Their 2019 paper in Circulation: Arrhythmia and Electrophysiology proposed something I find genuinely compelling, that the discrepancy between ECG-predicted age and actual age might work as a physiological measure of health.

Here's the part I want you to hold onto, though. These are population-level associations. A 2% bump in mortality risk per year of retinal age gap is real and it reproduces across tens of thousands of people. It is not a prophecy about you specifically. I've watched patients treat a biological age score as either a death sentence or a permission slip, and both readings miss what the number is.

The Trial Where an Algorithm Actually Changed What Happened to Patients

Prediction studies are relatively easy to publish. Trials testing whether the prediction improves care are difficult, expensive, and rare. Which is why the EAGLE trial matters more than nearly anything else in this space.

Yao and colleagues, again at Mayo, published results in Nature Medicine in 2021 from a pragmatic cluster-randomized trial. They took an AI tool that reads a standard ECG and flags patients likely to have a weak heart pump, specifically a low ejection fraction, a treatable condition that routinely goes undiagnosed until someone ends up in the hospital short of breath. Then they randomized 120 primary care teams across 45 clinics and hospitals. Half the clinicians saw the AI result. Half practiced as usual. In total, 22,641 adults with no known heart failure had an ECG as part of ordinary care.

The intervention worked. New diagnoses of low ejection fraction within 90 days rose from 1.6% under usual care to 2.1% when the AI result was available. Among the patients the algorithm actually flagged as high risk, diagnosis rates climbed from 14.5% to 19.5%. The full EAGLE trial results are published and open.

Those numbers look small on the page. They aren't. A half-percentage-point absolute increase across a primary care population means a meaningful number of people started treatment for a serious, progressive, treatable disease earlier than they otherwise would have. And here's the detail I appreciate most: echocardiogram use barely budged in the overall cohort. This wasn't an algorithm generating a mountain of expensive downstream testing to justify itself.

That's the bar I'd like this field held to. Not "our model predicted mortality," but "clinicians using our model found more disease without ordering more scans."

Where This Actually Shows Up in a Longevity Practice

Personalization is the word everyone reaches for, and it's been drained of most of its meaning. What it should mean is straightforward. Two patients walk in with identical complaints and leave with different plans, because their underlying biology is different and we have data showing how.

Machine learning helps with that in a few concrete, undramatic ways. Pattern recognition across a wide lab panel can surface combinations a human skimming a results page tends to slide past, especially when several values sit inside the reference range individually but are drifting together. Continuous data from a wearable gives a picture of sleep and cardiovascular fitness that no single office visit can. Risk models built on large cohorts help me tell a patient how much a given intervention is likely to move their particular needle instead of the average person's.

What none of it does is decide anything. When I'm evaluating whether someone is a reasonable candidate for longevity medicine and geroprotective medications, algorithmic output is one input among many, sitting alongside history, labs, goals, and how the person describes their own energy and function. The same holds when we're working through hormone optimization, where the numbers and the symptoms frequently disagree and a human being has to reconcile them.

In my practice, the most useful thing a biological age estimate has done is motivational. I had a patient in his early fifties who had shrugged off every conversation about his metabolic labs for two solid years. He got a score suggesting his cardiovascular system was tracking closer to sixty, and something finally clicked. He started training seriously, fixed his sleep, and his labs followed. That's a legitimate use of the technology and I'm not going to be snobbish about it. But notice what did the work. The intervention was resistance training and eight hours of sleep. The test was just the thing that got his attention.

What I'm Still Skeptical About

Several things, and I'd rather say them plainly than have you discover them after you've paid.

• Most consumer biological age tests are not the tools in these papers. The UK Biobank retinal model and the Mayo ECG networks were built and validated on enormous, deeply characterized cohorts. A product that mails you a cheek swab is a different animal wearing the same vocabulary.
• Reproducibility is uneven. Some biological age tests return meaningfully different answers if you take them twice in one week. Ask any vendor about test-retest reliability. The good ones will answer without flinching.
• Training populations matter. A model built largely on one demographic can perform worse on people outside it, and that gap doesn't always make it into the marketing copy.
• Association is not causation, and the distance between them here is enormous. Nothing in these studies demonstrates that shrinking your retinal age gap extends your life. It shows the two move together.

None of that makes the technology worthless. It makes it a tool with a defined range, like every other tool in medicine. And the question I'd want answered before anyone spends money is whether the result would change something they actually do. If you already know you sleep six hours a night, haven't lifted anything heavy since college, and skipped your last two rounds of labs, no algorithm is going to hand you a more actionable insight than that. Fix the obvious first. The obvious is where almost all of the return lives.

Machine learning is going to keep working its way into clinical practice, and on balance that's good news. The EAGLE trial is a preview of where it belongs: quiet, narrow tools that help a physician catch something earlier, embedded in care a person still directs. What I'd stay skeptical of is the version where a score replaces the conversation.

At Magnolia Functional Wellness in Southlake, I'm glad to talk through whether any of this has a place in your plan, including the perfectly reasonable answer that it doesn't yet. The fundamentals still do most of the heavy lifting. But if you want to understand what your data actually says about your body, that's a conversation worth having with someone who will also tell you when a number doesn't mean much.