Agentic AI for Scaling Diagnosis and Care in Neurodegenerative Disease

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Agentic AI for Scaling Diagnosis and Care in Neurodegenerative Disease".

Jane: The paper was written by Andrew G. Breithaupt, Michael Weiner, Alice Tang, Katherine L. Possin, Marina Sirota et al. from Goizueta Brain Health Institute, Emory University and Department of Radiology and Biomedical Imaging, University of California, San Francisco and School of Medicine, University of California, San Francisco and Bakar Computational Health Sciences Institute, University of California, San Francisco and Memory and Aging Center, Department of Neurology, University of California, San Francisco and Department of Neurology, Weill Institute for Neuroscience, University of California, San Francisco and NSF AI Institute for Advances in Optimization (AI4OPT), Georgia Institute of Technology and Department of Neuroscience, University of California, Berkeley and Division of Biostatistics, University of California, Berkeley and Center for Intelligent Imaging (ci2), Department of Radiology and Biomedical Imaging, University of California, San Francisco and Department of Psychiatry and Behavioral Sciences, University of California, San Francisco.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title and Authors: Tom: Welcome back to the show, everyone. Today we're looking at a paper that's got the whole team buzzing — it's called "Agentic AI for Scaling Diagnosis and Care in Neurodegenerative Disease." Jane, this one feels personal, doesn't it?

Jane: It really does, Tom. This is about Alzheimer's and related dementias, and the paper opens with some sobering numbers. The U.S. is facing a massive shortage of neurologists, and by two thousand sixty we're projected to see a million new dementia cases a year. Meanwhile, over half of dementia diagnoses in primary care are delayed until moderate or advanced stages.

Tom: And the wait times are brutal. The paper cites projections that by two thousand twenty-seven the average wait to see a dementia specialist could exceed forty months. Rural areas face even longer delays — three times worse.

Jane: Right, and that's the crisis this paper is trying to address. The authors come from UCSF, Emory, Berkeley, Georgia Tech — a real who's who in dementia research and AI. The lead authors include people from the Memory and Aging Center at UCSF, which is one of the premier institutions for this kind of work.

Tom: So what's their big idea? They're proposing something called agentic AI systems — and I know that sounds like jargon, but it's actually pretty intuitive once you break it down.

Jane: Exactly. Think of it this way: a regular AI chatbot can answer questions, but an agentic AI system can actually do things. It can pull data from electronic health records, run analyses, search medical literature, and coordinate all those steps to help a clinician make a diagnosis. It's like having a really smart assistant who doesn't just talk — they act.

Tom: And the key word here is "assist." The paper is very careful to say this is about augmenting clinicians, not replacing them. They call it keeping the human in the loop, which matters a lot when you're dealing with vulnerable patients who may have declining decision-making capacity.

Jane: That's the ethical backbone of the whole paper. They're proposing a six-phase roadmap — starting with standardized data collection, then building AI decision support, integrating it into clinical workflows, validating it rigorously, enabling continuous learning, and wrapping it all in ethics and risk management.

Tom: Six phases. That's ambitious. But before we get into the weeds of the roadmap, I want to ask Lu — you've been quiet. What's your first reaction to the title alone?

Lu: Honestly, Tom, I think the title undersells it. "Scaling diagnosis and care" sounds incremental, but this is really about rethinking how specialty medicine works. The authors are saying we can't train enough neurologists fast enough, so we need to build systems that multiply the expertise we already have. That's a fundamental shift in how we deliver healthcare.

Tom: Lu's right — this isn't just another AI paper. It's a blueprint for changing the system. And we're going to dig into that blueprint over the next few segments. Stay with us.

Summary of the Paper: Jane: So we're back with "Agentic AI for Scaling Diagnosis and Care in Neurodegenerative Disease," and I want to walk through what the paper actually proposes. Tom, you mentioned the six phases — let's talk about the first two, because they set the foundation.

Tom: Phase one is data collection, and the paper makes a really practical point: before any AI can help, you need good data. But right now, patient histories are collected inconsistently — one doctor asks different questions than another, and the notes are all over the place. So they're proposing AI-powered tools that standardize this from the start.

Jane: And they've got some concrete examples. Voice-based conversational agents that can take a patient history over the phone — imagine an AI that interviews a patient about their memory problems, their sleep, their mood, and does it in a way that's empathetic and adapts to the person's education level or language. That's not science fiction; the paper says these systems already exist and are being tested.

Lu: What excites me is the integration piece. They're talking about digital cognitive assessments that capture trial-level data — response times, error patterns — not just a summary score. That's a goldmine for AI. A traditional paper test gives you one number; a digital test gives you a rich dataset that can reveal subtle patterns.

Tom: And then phase two is decision support. This is where the AI actually helps interpret all that data. The paper describes a system that can look at a patient's history, cognitive testing, brain MRI, and blood biomarkers, and then generate a differential diagnosis with explanations the clinician can verify.

Jane: That's the part I find most compelling — the explanations. The authors emphasize that clinicians won't trust a black box that just says "this patient has Alzheimer's." They need to see the reasoning. So the system is designed to mimic how a clinician thinks, presenting evidence in familiar clinical terms.

Meng: Can I jump in here? From an engineering standpoint, the multimodal integration is the hard part. You're combining free-text clinical notes, structured lab results, MRI images, and speech patterns from the patient interview. Each of those is a different data type that needs different processing. The paper actually describes a system that converts imaging features into textual summaries that language models can reason over — that's a clever workaround.

Lu: And they cite real examples. There's a system called MAI-DxO from Microsoft that does sequential diagnostic reasoning, and Google's multimodal AMIE for consultations. So this isn't purely theoretical — the building blocks exist.

Tom: But here's the thing that struck me — the paper is honest about the gap between what works in a research setting and what works in a real clinic. They mention that most generative AI diagnostic studies use simulated data, not real patients. So there's a big validation gap they're trying to close.

Jane: And that's exactly what phase four is about — validation and monitoring. But before we get there, we need to talk about how this actually fits into a clinician's day. That's the workflow piece, and I think it's where the paper gets really interesting. We'll dig into that next.

Improvements Suggested by the Paper: Tom: Welcome back. We're still on "Agentic AI for Scaling Diagnosis and Care in Neurodegenerative Disease," and Jane just teed up the workflow question. This is where the paper makes some really concrete suggestions about improving how clinicians work.

Jane: Right, and the key insight is that this system should save time, not add to the burden. The paper talks about using AI to collect the patient history before the visit even happens — over the phone or through a conversational agent. Then when the patient sees the doctor, the history is already summarized, the cognitive assessment is already done, and the doctor can spend the visit actually talking with the patient.

Tom: That's a radical shift. Instead of the doctor spending twenty minutes typing notes while the patient talks, they can focus on building rapport and making shared decisions. The paper calls this "shifting clinic visit time from data collection to meaningful tasks."

Meng: But I want to push back on something. The paper mentions electronic consults — e-consults — as a near-term implementation pathway. That's where a primary care doctor sends a case to a specialist electronically instead of referring the patient. The idea is that the AI system provides enough data that the specialist can answer more cases remotely. But that only works if the AI is trustworthy enough for the specialist to rely on it.

Jane: That's a fair point, Meng. And the paper addresses it through their validation framework — they call it FAVES, which stands for Fair, Appropriate, Valid, Effective, and Safe. They're proposing that models be tested against diverse, specialist-confirmed cases, not just neuropathology, because neuropathology cases tend to come from affluent, well-educated patients.

Lu: The continuous learning piece is what gets me excited. The paper describes a system where specialists use the AI to review cases, and every time they do, they're implicitly validating or correcting the AI. Over time, the system learns from that feedback. It's like the AI gets better every time a specialist uses it, and that improvement benefits every other clinician on the system.

Tom: That's the flywheel effect — the more it's used, the smarter it gets. But they're careful to say that learning has to be supervised. You can't just let the AI update itself based on any case. They emphasize careful case selection and human oversight to prevent what they call "model degradation."

Meng: And there's an economic angle here too. The paper acknowledges that the return on investment is uncertain. But they point out that ambient scribes — AI tools that automatically document visits — have already been adopted widely because they reduce clinician burnout. So there's precedent for healthcare systems investing in AI even without clear reimbursement.

Lalam: If I may add a perspective — the improvements here go beyond efficiency. This paper is about equity. By making specialist-level assessment available in primary care and rural settings, you're addressing the fact that minority populations face greater diagnostic delays. The authors explicitly call out the need to mitigate bias in these systems and ensure they work across languages and educational backgrounds.

Jane: That's a crucial point, Lalam. The paper isn't just about making a fancy tool — it's about making sure the tool works for everyone, not just the people who can already access good care. And that brings us to the ethics and risk management piece, which is where we'll wrap up.

Conclusion: Tom: We've covered a lot of ground on "Agentic AI for Scaling Diagnosis and Care in Neurodegenerative Disease." Let's pull it together. Jane, what's the big picture?

Jane: The big picture is that we have a crisis — not enough neurologists, too many dementia patients, and diagnoses coming too late. This paper proposes a six-phase roadmap to build AI systems that help clinicians work at specialist level, even when they're not specialists. It starts with standardized data collection, moves through AI decision support, integrates into workflows, validates rigorously, learns continuously, and wraps everything in ethics.

Tom: And the through-line is human oversight. Every phase keeps clinicians in the loop. The AI collects data, but a clinician reviews it. The AI suggests a diagnosis, but a clinician verifies it. The AI learns from cases, but specialists curate which cases it learns from.

Lu: What I'll remember is the vision of a continuously learning healthcare system. This isn't a static tool — it's a system that improves with every patient encounter and incorporates the latest research in real time. That's a fundamental shift from how medicine works today.

Meng: And from a practical standpoint, the near-term wins are real. E-consults, ambient scribes, digital cognitive assessments — these exist today. The paper gives a realistic path from where we are to where we need to be.

Lalam: The cultural impact is significant as well. This paper models how AI can serve vulnerable populations with dignity — respecting patient autonomy, ensuring transparency, and building accountability mechanisms. It sets a standard for how medical AI should be developed, not just for dementia but for all of medicine.

Tom: Well said. We've covered the title, the summary, the improvements, and the implications of "Agentic AI for Scaling Diagnosis and Care in Neurodegenerative Disease." It's a roadmap paper, but it's grounded in real systems and real challenges. Thanks to Lu, Meng, and Lalam for joining the conversation.

Jane: And thanks to our listeners. This is one of those papers that could genuinely change how we deliver care to millions of people. We'll be back soon with the next paper — until then, take care.

Tom: Goodbye, everyone.