For the last five years, every Alzheimer’s conference has featured at least one talk on artificial intelligence and risk prediction. The promise has always been the same: combine blood tests, imaging, cognitive scores, and clinical data, and let a trained model do what a doctor can’t do by hand — integrate a lot of different inputs and produce a useful risk estimate. The challenge has always been the same too. The AI models got trained on neat, tidy research data. Real-world clinical populations are messy. The models rarely worked in the real world the way they worked in papers.
The paper that closed the gap
Bülbül and colleagues in the Journal of Clinical Medicine (Bülbül 2026) built and tested a real-world AI model that combines blood tests, imaging, and clinical data that are already part of routine care. They ran it on real clinical data, not on a pristine research cohort. It was not as flashy as the research-cohort numbers, but it worked well enough to change which tests get ordered, for which patients, and when.
The distinction matters. Prior AI-in-Alzheimer’s papers routinely reported impressive performance in idealized research settings. Those results rarely replicated when clinics tried to use them on ordinary patients, because the research data had been cleaned and enriched in ways that real-world data isn’t. Bülbül’s paper is important precisely because it was tested in the environment where it has to work.
What this actually changes
Picture two patients walking into a memory clinic.
A 68-year-old with subjective memory complaints, no family history, a clean cognitive screening test, controlled blood pressure, normal blood work. A model with all of that as input says low risk. Expensive blood tests and PET scans can wait. The patient saves money, avoids an anxiety-inducing workup, and gets a concrete reassurance plan.
A 62-year-old with early cognitive slippage, APOE ε4 carrier status, insulin resistance, and subtle changes on a routine MRI. The model says high risk, so the blood tests get drawn this week rather than six months later after another round of watchful waiting.
Both decisions are better: the first saves healthcare resources and spares the patient an unnecessary workup, while the second catches a decline trajectory when there’s still time to modify it. That is the core promise of precision diagnostics, and in 2026 a real-world pipeline exists to deliver on it.
The unsolved question
The question is no longer whether AI can do this work in memory care. It’s whether individual clinics are set up to run it. The data, the hardware, the informatics support, and the clinician champion who will pilot it are each non-trivial. Most clinics today can’t do all four. The ones that solve the operational problems will be about 18 months ahead of the ones that don’t.
We are building that pipeline now at Rezilir. The infrastructure we put in place for the EVANTHEA trial has given us an organized clinical dataset of 250-plus biomarkers that a model can actually learn from, without the heroic data-cleaning effort most memory clinics would face. We also have developed an ongoing, growing research database that keeps contradictory findings visible instead of smoothing them away, and an IRB-approved registry so we can go deeper into real-world data with other clinics.
That infrastructure, not the AI algorithm itself, is probably the hardest part of this whole story.
What this means for you
In the short term, AI in memory care is a behind-the-scenes tool that helps doctors decide who needs what test, and when. You won’t necessarily see the algorithm. You’ll see its effects: faster triage to the right workup if you’re at higher risk, and less pressure for expensive tests if your risk profile says they’re not warranted yet.
Over the next few years, expect this kind of tool to become standard in well-equipped memory clinics. It will sharpen clinical judgment, not replace it.
Where this series lands
Across seven parts and roughly twenty papers, one thesis keeps showing up.
Mild cognitive impairment is not one disease. It is a meeting point for several underlying mechanisms: amyloid in a genetically-defined minority, metabolic dysfunction in the majority, chronic inflammation driven by infection or environmental exposure in a substantial fraction, hormonal loss in women who miss the timing window, and overlapping combinations in most real people who come in for help.
Treating MCI as one disease, with one drug or one diet or one supplement stack, is why so many of the big trials underperform their headlines. Treating it as a phenotyped individual — a person with their own specific risk profile and their own specific levers — is what we do every day at Rezilir, and what we continue to see produce outcomes in our clinic and in the EVANTHEA trial.
We will continue to do the following in our precision medicine practice as the evidence for this strengthens.
- APOE genetic testing on every new intake with cognitive concerns. Will expand this to further genomic testing if there are real cognitive deficits.
- Plasma p-tau217, GFAP, and NfL as the first tier of blood biomarkers.
- Environmental screening (chemicals, metals, biotoxins) in patients with cognitive concerns especially APOE ε4 carriers.
- HSV-1, HSV-2, and broader chronic-infection testing when the clinical picture supports it, with treatment reserved for the right individuals.
- Metabolic phenotyping before advancing any protocols.
- A graded exercise prescription chosen for what you’ll actually do, plus the intensity needed to produce a training effect.
- Transdermal bioidentical hormone therapy, for women, in the right candidate.
- Tirzepatide over semaglutide in a cognitively-at-risk type 2 diabetic.
- An anti-amyloid antibody conversation for the right APOE-stratified, biomarker-confirmed patient who still wants that treatment
- Post-COVID cognitive symptoms treated as a chronic inflammatory condition, with a six-month follow-up regardless of how someone feels at the first visit.
- AI-assisted risk review running in the background of every intake, pointing us toward the patients who need more workup and away from the ones who do not.
The point is that none of this is one-size-fits-all.
References — Part 7
- Bülbül NG, Baytaş İM, Kavalcı E, Karasu E, Okcu Korkmaz BC, Belen BG, Musaoğlu İS, Övüt AR, Arslanoğlu NE, Urhan M, Mutlu H, Özdağ MF. Real-World Multimodal Machine Learning for Risk Enrichment Across the Alzheimer’s Disease Spectrum. Journal of Clinical Medicine. 2026;15(6):2250. doi: 10.3390/jcm15062250
Taking this from reading to doing
If any of this applies to your situation and you want to go deeper, the Rezilir protocols — our Brain Stack, the precision biomarker panels, the supplement pairings matched to your specific picture — are available through our Fullscript dispensary, dosed and timed the way we use them in the clinic. Review them here.
This series is educational, not medical advice. Anything you take from it should be reviewed with your own clinician, especially if you have a diagnosed condition, take medications, or are pregnant or nursing.
