The Last Twenty Feet: Why Alzheimer's AI Stalls at the Clinic Door

A peer-reviewed narrative review maps six concrete blocks between validated Alzheimer's algorithms and the exam room, from unrepresentative training data to tools that cannot explain themselves. It is a synthesis of existing literature, not new patient data.

Imagine the twenty feet between a hospital's imaging server and the room where a family is told what the scan means. In press releases, that distance does not exist: an algorithm reads the MRI, flags the hippocampal atrophy, and the clinician nods along. In practice, that short walk is where most Alzheimer's artificial intelligence quietly dies. A narrative review published online on 18 December 2025 in Annals of Medicine and Surgery, by Muhammad Umer Suleman and colleagues at Ayub Medical College in Abbottabad, Pakistan, and BRAC in Dhaka, Bangladesh, sets out to name exactly what is in the way. It is a synthesis of published literature rather than a study of patients, and the authors say so plainly. But it is an unusually specific map of a problem families feel without being able to describe: the tools keep being announced, and they keep not arriving.

The review counts six barriers. Data sets that are small and unrepresentative. Algorithms that cannot explain their own reasoning. Disruption to the way clinics already run. Regulatory pathways that remain unclear. Implementation costs and infrastructure demands. And the risk, running under all of it, that the tools widen existing health disparities rather than closing them. Two of the figures the authors cite are worth sitting with. Citing prior published work, they report that AI models developed in the United States were 12 to 25 percent less accurate when tested on Asian and African populations, a gap they attribute to genetic, environmental and imaging differences rather than to any single flaw in the code. They also report, again citing prior work, that fewer than 8 percent of AI tools cleared by the US Food and Drug Administration provide any explanation of how they reached their decisions. Both numbers come to the review secondhand, from other papers, and should be read as such.

That second figure is the one that turns a technical problem into a human one. A neurologist who is handed a probability without a rationale is being asked to put their name, and their liability, on a black box. The review notes that radiologists and neurologists tend to want visible reasoning, heatmaps or highlighted regions, before they will act on a suggestion, and that opacity raises both liability and ethical concerns. Surveys in radiology summarised in the review found that clinicians' lack of knowledge about AI, lack of trust in it, and fear of being automated out of a role were themselves major barriers to adoption. Allied health professionals reported limited training and similar job anxiety. None of that appears in an accuracy metric. All of it determines whether a tool gets used on a Tuesday afternoon.

The generalisation problem has been measured more formally by some of the same authors. In a separate systematic review published in Annals of Medicine and Surgery in October 2025, registered with PROSPERO and reported to PRISMA standards, the team searched PubMed and Embase for imaging AI studies published between January 2022 and June 2025 that reported both internal and external validation. They found 342 records. After removing duplicates and screening, exactly six studies qualified. Six. The scarcity is the finding: the great majority of published radiology AI has simply never been tested outside the hospital that built it, which means the confident accuracy figures in the abstracts describe a performance that may not survive a change of scanner, let alone a change of continent.

The review's proposed fixes are the mirror image of its barriers: large collaborative data repositories, better methods for making model reasoning transparent, clinician education programmes, clear regulatory pathways, and investment in scalable infrastructure. It also points to federated learning, which trains models across institutions without moving identifiable patient data, and reports one study in which the approach improved detection performance, measured by area under the curve, by 11 percent across worldwide cohorts. That is a single cited result, not a settled conclusion. What the authors are describing is a direction of travel, not a delivered solution, and the honest reading of their paper is that the toolkit remains in the corridor.

For a family waiting on a diagnosis, the practical meaning is modest and worth knowing: if an AI tool is not part of your assessment, that is usually not because your clinic is behind. It is because most of these systems have not yet been shown to work outside the place that built them, and because a tool that cannot explain itself is hard for any clinician to stand behind. Questions about what is and is not being used in your own care belong with the clinician making the diagnosis.