Seeing It Coming: AI, Early Dementia Detection, and the Discipline of Doubt

Most families I know have been touched by dementia in one way or another.

My mother was diagnosed with rapid-onset dementia. Within months, she went from a fully functioning professional — a volunteer at her local botanical garden, an adored grandmother — to someone who could barely recognize the people she loved. Including me, her only child. It was devastating, and it was fast, and by the time the diagnosis had a name, her cognitive function had already fallen off a cliff.

I carry my own version of this fear. A fall off a roof left me with a severe traumatic brain injury, and since then I've had to ask myself, more than once, whether the cognitive friction I feel some days is early dementia announcing itself. I've done the neuropsych evaluations. I've had the MRI and PET scans. Every test says the same thing: no signs of early dementia. I believe them. But I understand, from the inside, why someone would want a faster answer than a battery of tests that only speaks up once the disease is already loud.

That question — could AI tell us sooner? — is where this piece starts. It can't be where it ends. The same instinct that sent me to neuropsych evaluations instead of a hunch is the instinct this technology demands of all of us: don't just ask what AI could do. Ask what it has actually been shown to do.

Part One: The Promise

The scale of the problem is not in dispute. An estimated 55 million people worldwide are currently living with dementia, and that number is projected to nearly triple by mid-century. What is in dispute — what has always been in dispute — is timing. People are diagnosed an average of 3.5 years after symptoms are first noticed, and 4.1 years for those with early-onset dementia, according to a University College London meta-analysis published last year. That is not a rounding error. That is years of a family adjusting to a stranger wearing a familiar face, without a name for what is happening or a plan for what comes next. alzintnews-medical

This is the gap AI vision research is aiming at, and the early signals are genuinely interesting — not because any of it is ready for your doctor's office tomorrow, but because the biological logic holds up. Alzheimer's disease appears to leave fingerprints on how we move our eyes and our faces long before it shows up on a cognitive test. A growing body of research — from saccadic eye-movement studies to deep convolutional neural networks trained on eye-tracking data — has documented measurable differences in visual scanning and gaze patterns between people with Alzheimer's, mild cognitive impairment, and healthy controls. A recent systematic review pooling these studies found that AI models analyzing facial expressions demonstrate meaningful accuracy in detecting neurocognitive disorders in older adults, with one deep-learning model reaching 90.63% accuracy identifying mild cognitive impairment from full-face video. Cross-Enhanced Multimodal Fusion of Eye-Tracking and Facial Features for Alzheimer's Disease Diagnosis +2

The applied versions of this research are starting to take shape, too. At Texas A&M, researchers are building an AI-powered "digital human" that combines screening questions with facial expression analysis and biometric monitoring to catch subtle early signals like apathy — one of the quieter, earlier indicators of dementia that traditional screening tends to miss. And this isn't confined to faces and eyes: at Stanford's Alzheimer's Disease Research Center, investigators are using computer vision to automatically extract gait speed, posture, and movement features from simple video recordings, turning them into the same clinical mobility scores a physical therapist would generate by hand — because gait changes, too, can be an early tell. tamunih

Put simply: the idea that a camera and a few minutes of conversation could flag risk during a routine visit, years before a formal diagnosis, is not science fiction. It's an active, multi-institution research agenda with encouraging preliminary numbers behind it.

And that's exactly where I have to stop myself. Because "encouraging preliminary numbers" is not the same sentence as "ready to screen your mother." Conflating the two is how families get hurt.

The Turn

Here's the thing about the tests that cleared me. Nobody handed me a headline and asked me to trust it. I sat for hours of neuropsych evaluation. I had my brain imaged twice, with two different technologies, by people whose entire job is knowing what those images do and don't prove. The reassurance I got wasn't reassurance — it was evidence, produced by a process built to be skeptical of itself.

That's the standard AI health claims deserve too. Not because the researchers doing this work are being dishonest — the studies above are legitimate, peer-reviewed, and represent real scientific progress. But because the distance between "a peer-reviewed pilot study found promising accuracy in 40 participants at one clinic" and "a tool that reliably screens your father for dementia" is enormous, and headlines routinely erase that distance. If we're going to let AI into the most vulnerable moments of our families' lives, we owe it the same scrutiny we'd demand of any clinician walking into that room.

Part Two: The Discipline of Doubt

So how should a patient, a caregiver, or a clinician actually evaluate a new AI health claim? Researchers at Stanford's Institute for Human-Centered AI recently proposed a framework that cuts through the noise with three plain questions: What exactly is being claimed? What was actually tested? And do the two match? It sounds almost too simple. In practice, it's the question most AI health coverage never bothers to ask. stanford

I'd add four practical checkpoints underneath it, drawn from how clinical AI is actually supposed to be vetted:

1. Was it tested on people, or on a dataset? A model that classifies pre-collected images in a lab is a different animal from a model tested prospectively on new patients walking through a clinic door. Retrospective accuracy is a starting point, not a verdict.

2. How big, and how diverse, was the sample? Dementia doesn't present identically across ages, ethnicities, or comorbidities like my own traumatic brain injury. A tool validated on fifty participants at one academic medical center has not yet earned the right to generalize to fifty million people worldwide.

3. Has it been independently replicated, and is the evidence peer-reviewed? Systems built on outdated, unvetted, or non-peer-reviewed inputs can introduce real risk and mislead the clinicians relying on them. One promising paper is a hypothesis. Several independent teams reaching the same result, using different populations, is evidence. ebsco

4. Is there a governance framework, or just a product? There is still no single, agreed-upon standard for evaluating AI-based health technologies the way there is for a new drug or device — which means, for now, the burden of asking "who validated this, and how, and how often is it re-checked?" falls on us. nih

Applying the Framework to This Very Article

In the spirit of the discipline I'm asking for, let me turn it on the research I just cited.

Most of the eye-tracking and facial-expression studies above are exactly what good early-stage science looks like: promising, peer-reviewed, and small. Sample sizes in the dozens to low hundreds. Single-institution. Largely retrospective. None of the tools described above are, to my knowledge, FDA-cleared diagnostic devices — they are research prototypes and pilot studies, several years and several validation rounds away from a clinician's exam room. The 90.63% accuracy figure I cited above is real, and it's also a pooled result from a meta-analysis of heterogeneous studies — worth taking seriously, not worth mistaking for a green light.

That doesn't diminish the promise. It defines it honestly. This is what "the simplest version that would still be amazing" actually looks like in year one: not a finished product, but a research trajectory worth funding, watching, and — eventually, if the evidence holds — bringing to primary care.

Where This Leaves Us

My mother didn't get the years she deserved because the system that was supposed to catch her illness caught it too late. If AI can close even part of that 3.5-to-4.1-year gap, responsibly, on evidence that earns the trust it's asking for, that is worth fighting for. Families with a history of dementia would rightly demand access to a screening tool like this the moment it's proven. So would I.

But sovereignty over your own health — patient or clinician — was never about accepting the newest tool because it promised the most. It was about being equipped to ask it the same hard questions I asked of my own scans. What exactly are you claiming? What did you actually test? Do the two match?

That question, asked consistently, is not skepticism for its own sake. It's how we get to keep the promise without getting fooled by the pitch.