AI in Medicine Update — Oct 1, 2026

This AudioScholar briefing is generated by artificial intelligence for healthcare professionals and trainees. It is not medical advice.

Welcome to your weekly update in AI in Medicine. The papers in this period circle one question from four directions: what happens when these models stop being benchmark performers and start sitting inside a clinical workflow. We have a hematology decision-support agent taken all the way to a prospective silent trial, three foundation models reading tissue and transcriptomes for treatment selection, two consensus documents from radiology that disagree with each other about whether artificial intelligence actually makes radiologists faster, and two sobering papers on what breaks underneath — adversarial brittleness in the frontier models, and privacy risk that falls hardest on underrepresented patients.

Start with large language models placed directly into the diagnostic conversation, because this is where the evidence moved furthest. In Nature Medicine, Zoller and colleagues built HemaGuide, a locally deployable agent for hematological malignancies that turns unstructured clinical documents into structured case representations, routes each case to a guideline, advanced or molecular reasoning mode, and grounds its recommendation in disease-specific flowcharts plus a memory of more than two thousand real tumor board cases. The clinical question was whether subspecialty tumor board reasoning can be reproduced where that deliberation isn't available. In expert-blinded benchmarking on forty-five high-complexity cases, concordance with actual board decisions improved substantially, and an ablation across eleven layers showed no single component carried the gain — it was routing-dependent. More importantly, external validation on five hundred and fifty-five independent cases from a second academic center reached roughly eighty-two percent concordance across forty-seven disease entities, and a prospective one-month silent trial on sixty-four consecutive unselected cases landed in the same range. Hallucinations occurred in two of six hundred and sixty-four evaluated cases. Median latency was thirty-nine seconds on commodity hardware, against the hours a manual molecular board workflow typically consumes. The limitations matter: both centers were academic, the simulated practice study — in which agent-assisted residents approached senior-level concordance — is a simulation and not patient outcomes, and concordance with a tumor board is agreement, not correctness. Still, this is among the first LLM decision-support systems with external and prospective validation reported together.

Sitting alongside that, in Radiology, Hong and colleagues offer a practical account of generative artificial intelligence for chest radiograph reporting. This is an experience-based how-to rather than a trial, so it carries no effect estimate, and should be read as expert framing. Its central claim is a posture: the model is a draft assistant, never an autonomous reporter, with explicit safety checkpoints and human-in-the-loop review built into the interface. The authors are candid that hallucinated findings remain the known failure mode, and that multi-institutional evaluation and explainability are still outstanding. What's striking is how closely their framing matches what Zoller's team engineered — grounding, auditability, a human at the end.

That convergence makes the next pair uncomfortable. Also in Nature Medicine, Gu and colleagues stress-tested flagship frontier models — the GPT-5 and Gemini generation — against adversarial transformations of health benchmarks. The finding is genuinely unsettling: these systems can often produce the correct answer with key inputs removed, meaning they are exploiting benchmark structure rather than reasoning from the clinical data, and yet they can be derailed by trivial prompt alterations, fabricating convincing but flawed reasoning traces along the way. Using clinician-guided rubrics, the authors show popular health benchmarks vary widely in what they actually measure. This is an evaluation study, not a clinical trial, and it tests general-purpose models rather than grounded, retrieval-anchored agents — which is precisely the tension. Zoller's agent reported hallucination in roughly one case in three hundred; Gu's team reports pervasive brittleness in the underlying model class. These do not obviously reconcile. The plausible explanation is that case-grounding and guideline flowcharts constrain the model enough to suppress the failure mode, but no one has yet run adversarial stress tests against a deployed clinical agent. That experiment — perturbing the inputs of a system like HemaGuide and seeing whether concordance holds — is what would settle it, and until it exists, benchmark performance and deployment robustness should be treated as separate claims.

And in Nature, Knolle and colleagues add a second kind of breakage. Across a diverse range of medical datasets, they performed patient-level privacy audits using membership inference attacks, which ask whether a given individual's data trained a model. Aggregate privacy metrics looked close to random guessing — reassuring, apparently — while individual patients could be identified with near-perfect success. The number of highly exposed patients rose with model capacity, and exposure concentrated in underrepresented groups, whether stratified by disease status, self-reported race, insurance, sex or imaging protocol. The authors are careful that whether this disparity extends beyond membership inference attacks is unknown. The implication for practice is evidentiary rather than procedural: the privacy assurances currently attached to medical models are computed in aggregate, and this work indicates aggregate numbers can severely understate what an individual patient faces — with the burden falling on the groups already least well served.

Turning to tissue and molecules, three foundation models push biomarker work toward something usable. In Nature Medicine, Shen and colleagues present COMPASS, a pan-cancer model predicting checkpoint inhibitor response from bulk tumor transcriptomes, routed through forty-four biologically grounded immune concepts rather than an opaque embedding. Trained on just over ten thousand tumors across thirty-three cancer types and tested across sixteen clinical cohorts spanning seven cancers and six checkpoint inhibitors, it outperformed twenty-two competing methods on average, improving accuracy by roughly eight percent, and generalized to cancer types and treatments it had not been fine-tuned on. Patients it classified as responders lived substantially longer. The caveats are real — these are retrospective cohorts, bulk transcriptomes require sequencing infrastructure, and nothing here has been tested as a prospective treatment-selection tool. The authors position it for indication selection and trial stratification, which is the honest framing.

Also in Nature Medicine, Vorontsov and colleagues report PRISM2, a slide-level pathology model trained on two point three million whole-slide images paired with fourteen million question-answer pairs derived from seven hundred thousand pathology reports. Supervising on diagnostic language, rather than image labels alone, is the novelty. With prompt-based inference, it matched or exceeded the balanced accuracy of clinical-grade commercial products calibrated for cancer detection in prostate, breast and breast lymph node, and its embeddings did not statistically underperform prior foundation models on diagnostic, biomarker and survival benchmarks. These remain benchmark comparisons rather than prospective reads against practicing pathologists.

And in Nature, Wenckstern and colleagues describe Virtual Tissues, a spatial proteomics foundation model that handles the field's central nuisance — every study uses a different marker panel. From one pretrained backbone it performs cell segmentation, typing, niche annotation and patient stratification, including annotation across heterogeneous panels it has not seen. In triple-negative breast cancer, biomarkers derived from it predicted anti-PD-L1 chemo-immunotherapy response and stratified disease-free survival in an independent cohort, outperforming both existing biomarkers and current clinical stratification schemes. Two of these three papers converge on immunotherapy response prediction from entirely different data types — transcriptome and spatial protein imaging — which is encouraging, but neither has been compared head-to-head against the other, nor against PD-L1 immunohistochemistry in a prospective design.

Which brings us to capacity, where radiology is arguing with itself. Two papers in Radiology, both from the International Society for Strategic Studies in Radiology, address the same gap between imaging demand and workforce. Moy and colleagues issue a call to action — a consensus statement from the society's Dublin meeting — spanning image interpretation, capacity building, data governance and systemic digital transformation, and arguing that artificial intelligence can increase workflow efficiency and align radiology with payer and value-based care priorities. Bredella and colleagues, reviewing the same workforce imbalance, state plainly that although digital tools have been introduced to increase capacity, their effect on radiologist productivity has been mixed. That is a direct disagreement about the central premise, and neither document resolves it, because both are consensus and review rather than measurement. What would settle it is prospective productivity data from deployed systems — reads per hour, turnaround, error rates, burnout — which neither paper has. Bredella's review also notes that teleradiology and hybrid work improved access in rural and underserved regions while contributing to isolation and lost collegiality.

The physical version of that labor gap appears in Nature, where Liang and colleagues ask how close humanoid robots are to laparoscopic surgery. They built a humanoid teleoperation framework using general-purpose instruments and tested it on the bench, in dry-laboratory studies across experience levels, and in vivo in pigs. The honest answer is feasible but not ready — task performance remains below established purpose-built surgical platforms, and the authors enumerate precision, control and safety gaps that must close before any clinical deployment. No human subjects here, and the value is that it sets a measured baseline rather than a promise.

If you read only one paper from this period, make it the HemaGuide study in Nature Medicine. It is the first clinical decision-support agent to carry external multi-center validation and a prospective silent trial in the same report, and it reopens a question many had treated as closed — whether language-model decision support belongs only in benchmark papers, or whether a locally deployable, auditable version can hold its concordance in consecutive unselected cases.

Here is what this period adds up to. Grounded, narrowly scoped language-model agents now have the strongest clinical evidence in this space, but that evidence comes from academic centers and measures agreement with tumor boards rather than patient outcomes. Benchmark performance and robustness have come apart decisively — the adversarial work shows frontier models can score well while reasoning from the wrong things, so published benchmark numbers should no longer be read as deployment readiness. Foundation models for immunotherapy response are converging from independent data modalities and beating current stratification schemes retrospectively, which strengthens the case for prospective trials but does not yet support clinical use. Privacy assurance methods are now known to be measuring the wrong thing, with risk concentrated in underrepresented groups. And radiology's own consensus bodies disagree about whether these tools improve productivity at all — watch for prospective workflow data, because that is the missing evidence, not more consensus.

That's your AI in Medicine update for this period. Until next time.

This is an automated summary generated by artificial intelligence, which can make mistakes. Always review the original source materials.