AMD follow-up gets a more auditable AI assistant
Longitudinal care for age-related macular degeneration hinges on decisions made visit after visit: is the disease active, has treatment response changed, and should the anti-VEGF interval be shortened or extended? Those judgments are rarely made from a single scan. They depend on a moving target built from serial retinal imaging, prior treatment history, and the clinician’s ability to synthesize the pattern quickly enough to guide the next injection plan.
The problem is not simply one of speed. Manual review can miss details buried across visits, and large language models, while attractive for summarizing complex records, can produce fluent but weakly grounded recommendations. In retinal medicine, where a small change in fluid burden or structural anatomy may alter treatment frequency, an answer without traceable evidence is not enough.
Evidence grounding replaces free-floating summaries
To address that gap, investigators developed AMDagent, a multimodal, tool-augmented clinical agent designed for evidence-based decision support in longitudinal AMD follow-up. The system combines quantitative imaging analysis with guideline retrieval, then converts those inputs into structured outputs intended for clinician review. Instead of offering a standalone narrative, the agent links its conclusions to measurable retinal findings and to evidence that can be checked at the bedside or in the chart.
A single sentence captures the practical advance: the model was built to justify its reasoning, not merely state it.
This matters because auditability is central to clinical trust. When a model can point to the underlying imaging measurement and the associated guideline source, the clinician is not asked to accept a black box. The output becomes a draft decision-support layer, not an unchallengeable directive. For a disease like AMD, where anti-VEGF scheduling often evolves over months, traceable reasoning is more useful than polished prose.
20 curated cases tested whether the agent could keep pace with real follow-up logic
The proof-of-concept evaluation used 20 curated longitudinal AMD cases. Expert adjudication rated the agent clinically appropriate in 92.0% of predefined decision-support units. In parallel, 83.3% of audited major clinical claims were explicitly grounded in tool-derived quantitative measurements and/or retrieved guideline evidence.
Those two numbers tell a coherent story. The first suggests that the agent usually produced follow-up guidance judged suitable by experts. The second suggests that most major claims could be traced to a measurable or guideline-based source, which is precisely the feature needed when decisions are being documented across repeated visits. In retinal practice, that kind of traceability can reduce friction during sign-out, second review, and longitudinal treatment planning.
Serial imaging is where this approach should earn its keep. AMD management is inherently cumulative: the current visit only makes sense when set against prior scans, prior response to therapy, and the trajectory of disease activity. By pairing image-derived data with retrieval from guidelines, the agent attempts to reconstruct that cumulative logic in a format clinicians can inspect rather than simply trust.
Why transparency matters in anti-VEGF scheduling
Anti-VEGF treatment intervals are often adjusted based on subtle evidence of stability or recurrence. A missed or overcompressed interval can have downstream consequences for visual prognosis, while unnecessary treatment adds burden without clear benefit. Because the consequences of interval selection are operational as much as they are biological, any support tool must show how it reached the recommendation.
The logic behind evidence-grounded assistance is straightforward. Quantitative imaging measures can anchor the assessment of disease activity, while guideline retrieval provides the clinical frame for interpreting those measures. Together they can reduce the risk that a model will invent a plausible but unsupported explanation. In AMD follow-up, that is the difference between an answer that sounds reasonable and one that can be defended in a clinic note.
The authors positioned the system as clinician-facing decision support rather than autonomous management. That is the right boundary for a first-generation tool in a specialty where small interpretive errors can affect the timing of injections and follow-up visits. The study design reflects a preference for verification over automation, which is a safer operating principle in ophthalmology.
What the proof-of-concept does and does not establish
The evaluation was preliminary, and the sample was small. A 20-case curated set can show whether the framework behaves sensibly, but it cannot settle how the agent will perform across broader real-world populations, more heterogeneous imaging quality, or messier longitudinal records. Nor can it determine whether the tool shortens visits, improves adherence to treatment intervals, or changes visual outcomes.
Even so, the signal is clinically relevant. The combination of multimodal analysis and guideline retrieval addresses one of the central weaknesses of conventional LLM-based summarization: the tendency to generate output that is hard to audit. Here, the emphasis on explicit grounding moves the system closer to a workflow that retina specialists might actually inspect and use. Expert-rated appropriateness and grounded major claims are early markers of that usability.
The study also points to a deeper issue in digital medicine: usefulness is not the same as fluency. A model that can summarize longitudinal AMD data while exposing its evidence trail is more aligned with how clinicians think, document, and defend decisions. In a field where treatment decisions are often repeated and revised over time, the ability to connect a recommendation to both measurement and guideline evidence may matter as much as the recommendation itself.
Larger prospective testing will decide whether the agent belongs in routine retina care
The next step is larger prospective multi-center evaluation. That is necessary to test generalizability across settings, imaging platforms, and workflow patterns, and to determine whether the system retains its performance when confronted with the variability of everyday retina clinics. It will also be important to assess whether the evidence-grounded format meaningfully changes clinician confidence, documentation quality, and the consistency of interval decisions.
For now, AMDagent sits in a promising but early category: a clinician-facing assistant built to preserve the chain of evidence in a disease that depends on serial interpretation. Its main contribution is not a new therapeutic target or a new retinal biomarker. It is a method for making longitudinal decision support more transparent, more reviewable, and less dependent on opaque language alone.