AI in Medicine Update — Oct 31, 2026
Generated Nov 1, 2026 · 9:25
The latest research on this topic, summarized for clinicians.
If the audio fails to play, refresh the page to renew the link.
Get this every week in your podcast app — free.
New research episodes land in your feed automatically — listen on your commute.
Impact of Unrestricted Artificial Intelligence Scribe Deployment in a Single Health System on Clinician Workload and Productivity : A Quality Improvement Study.
Across 1.64 million encounters, AI scribe enrollment cut note-active time by 1.21 minutes, with larger gains among high users, but did not improve same-day or after-hours note closure or wRVUs.
Annals of Internal Medicine · 2026 · PubMed
Papers in this briefing
- 01
Impact of Unrestricted Artificial Intelligence Scribe Deployment in a Single Health System on Clinician Workload and Productivity : A Quality Improvement Study.
Across 1.64 million encounters, AI scribe enrollment cut note-active time by 1.21 minutes, with larger gains among high users, but did not improve same-day or after-hours note closure or wRVUs.
Annals of Internal Medicine · 2026
From the briefingWe start at the point of care, where AI tools are already in clinicians' hands and the question has shifted from whether they work in a demo to whether they change the working day. Booth and colleagues, writing in the Annals of Internal Medicine, examined unrestricted deployment of the Heidi ambient scribe across a fourteen-hospital Boston system [1]. The study covered more than one and a half million outpatient encounters from 902 enrolled clinicians. Enrollment was associated with an immediate drop in active note-writing time of a little over a minute per note. The benefit was concentrated in heavy users, whose reduction was closer to two and a half minutes. Enrolled clinicians who never used the tool saw essentially nothing.
If you read only one paper from this period, make it Booth and colleagues on the ambient scribe [1]. It reopens the question of what scribes are actually for. They appear to save time on the note without moving closure rates or productivity, which shifts the debate from adoption to measurement.
Written by AI from the published abstract, for healthcare professionals. Not medical advice; check the paper before acting on it.
- 02
A clinical multimodal vision language foundation model with fine-grain explainability.
A multimodal vision-language foundation model reached 0.895 average diagnostic accuracy across 10 imaging modalities with interpretable reasoning (4.25/5 human rating), though prospective clinical validation is still required before deployment.
Nature Medicine · 2026
From the briefingA very different point-of-care tool comes from Bai and colleagues in Nature Medicine [2]. They describe a multimodal vision-language foundation model trained on more than a million samples. It spans ten imaging modalities across dermatology, hepatology, ophthalmology, pathology, pulmonology and radiology. In internal testing it diagnosed with about ninety percent average accuracy, beating both generalist and specialist models. It also highlights the image regions behind its reasoning, and human raters scored those explanations at better than four out of five. External testing on seventeen further datasets held up reasonably well. The contrast with the scribe study is instructive. The scribe study has messy real-world data and no clinical outcomes, while this paper has clean benchmarks and no deployment at all. The authors say plainly that prospective studies of safety and clinical utility are needed before routine use. Explainability that pleases reviewers is not yet evidence that it helps clinicians catch the model's errors.
Written by AI from the published abstract, for healthcare professionals. Not medical advice; check the paper before acting on it.
- 03
Autonomous biomedical research with an artificial intelligence agent.
An LLM-based agent autonomously executed diverse biomedical tasks, including rare-disease diagnosis, causal gene prioritization and drug repurposing, without task-specific tuning, suggesting AI could augment researchers' workflows.
Science · 2026
From the briefingHuang and colleagues, in Science, introduce Biomni, a general-purpose biomedical agent [3]. It mines tools and protocols from thousands of publications and then writes and runs its own analysis code. On benchmarks it generalized across tasks including rare-disease diagnosis, causal gene prioritization and drug repurposing, without task-specific tuning. Case studies show it orchestrating lab instruments and generating testable protocols.
Written by AI from the published abstract, for healthcare professionals. Not medical advice; check the paper before acting on it.
- 04
The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development.
AI agents annotated outcomes from 55,984 trials and found drugs targeting cell type-specific genes were 48% likelier to reach market with 32% fewer adverse events, informing therapeutic development decisions.
Science · 2026
From the briefingAlso in Science, Zhang and colleagues describe the Virtual Biotech, a multi-agent system organized like a drug company with divisions for targets, safety and clinical development [4]. Its headline finding came from more than 37,000 agents annotating outcomes across roughly 56,000 trials. Drugs aimed at cell-type-specific genes were about half again as likely to reach market and carried fewer adverse events. The system also proposed a lung cancer strategy and offered possible explanations for a failed ulcerative colitis trial.
Written by AI from the published abstract, for healthcare professionals. Not medical advice; check the paper before acting on it.
- 05
From code to cure: Harnessing AI for antibiotic discovery and design.
AI now enables exploration of vast chemical spaces and design of synthesizable antibiotics, but emerging reporting standards and regulatory hurdles must be addressed to bring these agents to patients.
Science Translational Medicine · 2026
From the briefingKrishnan, Collins and colleagues, in Science Translational Medicine, review AI for designing small-molecule and peptide antibiotics [5]. As a review, it adds no new data. Its value lies in flagging that data standards, reporting norms and regulatory pathways lag well behind the design capability.
Written by AI from the published abstract, for healthcare professionals. Not medical advice; check the paper before acting on it.
- 06
How to build an AI-driven digital organism.
A proposed AI-driven digital organism would integrate multiscale foundation models to predict and simulate biology from molecules to individuals, offering a safer, cheaper platform for guiding experimentation and medicine.
Nature Medicine · 2026
From the briefingSong, Segal and Xing, in Nature Medicine, offer a perspective proposing an AI-driven digital organism [6]. This would be a set of linked foundation models simulating biology from molecules to whole individuals. It is a vision rather than a result. For clinicians, the practical reading of this whole cluster is that these tools may shape which drugs reach trials within years, but none has yet shown benefit in a patient.
Written by AI from the published abstract, for healthcare professionals. Not medical advice; check the paper before acting on it.
- 07
AI-based characterization of Alzheimer's disease phenotypes from population-scale single-cell data.
A graph neural framework identified about 1.5 million phenotype-associated cells in Alzheimer's brains, prioritizing microglial and astrocyte subpopulations and pathways, validated in external datasets, for potential precision diagnostics and therapeutics.
Nature Medicine · 2026
From the briefingCloser to the clinic again, two studies use machine learning to extract biomarkers from molecular data. He and colleagues, in Nature Medicine, applied a graph neural framework to more than six million nuclei from the prefrontal cortex of over a thousand brains [7]. Their aim was to link specific cell subpopulations to Alzheimer's phenotypes. They identified microglial subtypes tied to pathology and reactive astrocytes that may confer cognitive resilience. In cognitively impaired donors they found excitatory-inhibitory imbalance and mitochondrial dysfunction. They also found astrocyte signatures shared between Alzheimer's disease and depression. Findings replicated in external datasets, and the tool is openly available. But this is post-mortem, cross-sectional tissue work, so the associations generate targets rather than diagnostics.
Written by AI from the published abstract, for healthcare professionals. Not medical advice; check the paper before acting on it.
- 08
Biomarkers of nivolumab benefit in resectable non-small cell lung cancer.
In CheckMate 77T biomarker analysis, a machine-learning model ranked pre-surgical ctDNA clearance, non-N2 disease, pCR, squamous histology and nivolumab as top predictors of prolonged event-free survival.
Nature · 2026
From the briefingCascone and colleagues, in Nature, reported biomarker analyses from the CheckMate 77T trial of perioperative nivolumab in resectable non-small cell lung cancer [8]. Among patients on nivolumab whose circulating tumour DNA cleared before surgery, half achieved a pathologic complete response, compared with about one in eight on placebo. Patients who turned molecular-residual-disease positive during adjuvant treatment all relapsed. Tumours carrying KEAP1, STK11, CDKN2A or SMARCA4 alterations, often considered resistant to immunotherapy, appeared to benefit, with the risk of an event roughly halved. A machine-learning model ranked ctDNA clearance, non-N2 disease, complete response and squamous histology among the top predictors of event-free survival. The important caveat is that only about four in ten randomized patients had evaluable biomarkers, so these subgroup signals are hypothesis-generating. Notably, the AI here confirms markers clinicians already recognize rather than discovering new ones.
Written by AI from the published abstract, for healthcare professionals. Not medical advice; check the paper before acting on it.
The full briefing
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. This period runs on three threads. The first is the most immediate: a large real-world look at ambient AI scribes suggests they save documentation time, but less than the hype implies, and without lifting productivity. The second is a cluster of papers in Science and Nature Medicine describing AI agents that design experiments, mine tens of thousands of trials, and propose drugs. Their ambition is striking, but they have not yet been tested prospectively. The third is AI turning molecular data into candidate biomarkers, from Alzheimer's brains to resectable lung cancer, where the clinical signal is promising but still exploratory.
We start at the point of care, where AI tools are already in clinicians' hands and the question has shifted from whether they work in a demo to whether they change the working day. Booth and colleagues, writing in the Annals of Internal Medicine, examined unrestricted deployment of the Heidi ambient scribe across a fourteen-hospital Boston system [1]. The study covered more than one and a half million outpatient encounters from 902 enrolled clinicians. Enrollment was associated with an immediate drop in active note-writing time of a little over a minute per note. The benefit was concentrated in heavy users, whose reduction was closer to two and a half minutes. Enrolled clinicians who never used the tool saw essentially nothing.
The finding that complicates the story is that heavy users also produced notes roughly 170 characters longer. And there was no improvement in same-day note closure, in after-hours closure, or in work relative value units. Time came off the keyboard, but it did not visibly translate into finishing notes sooner or seeing more patients. The design is an interrupted time series, which is a quality improvement study rather than a trial. It carries real risk of unmeasured confounding, since clinicians who chose to use the tool heavily may differ from those who did not. The authors also could not tell AI-drafted text from clinician-written text after enrollment. The evidence supports a modest, use-dependent time saving. It does not yet support claims about burnout or throughput.
A very different point-of-care tool comes from Bai and colleagues in Nature Medicine [2]. They describe a multimodal vision-language foundation model trained on more than a million samples. It spans ten imaging modalities across dermatology, hepatology, ophthalmology, pathology, pulmonology and radiology. In internal testing it diagnosed with about ninety percent average accuracy, beating both generalist and specialist models. It also highlights the image regions behind its reasoning, and human raters scored those explanations at better than four out of five. External testing on seventeen further datasets held up reasonably well. The contrast with the scribe study is instructive. The scribe study has messy real-world data and no clinical outcomes, while this paper has clean benchmarks and no deployment at all. The authors say plainly that prospective studies of safety and clinical utility are needed before routine use. Explainability that pleases reviewers is not yet evidence that it helps clinicians catch the model's errors.
Moving upstream, a group of papers asks whether AI can do the research itself, from hypothesis to protocol. Huang and colleagues, in Science, introduce Biomni, a general-purpose biomedical agent [3]. It mines tools and protocols from thousands of publications and then writes and runs its own analysis code. On benchmarks it generalized across tasks including rare-disease diagnosis, causal gene prioritization and drug repurposing, without task-specific tuning. Case studies show it orchestrating lab instruments and generating testable protocols. Also in Science, Zhang and colleagues describe the Virtual Biotech, a multi-agent system organized like a drug company with divisions for targets, safety and clinical development [4]. Its headline finding came from more than 37,000 agents annotating outcomes across roughly 56,000 trials. Drugs aimed at cell-type-specific genes were about half again as likely to reach market and carried fewer adverse events. The system also proposed a lung cancer strategy and offered possible explanations for a failed ulcerative colitis trial.
There is a genuine tension between these two papers. Biomni is framed as autonomous. The Virtual Biotech authors deliberately describe their system as human-guided and stress transparency. Neither paper reports a prospective test in which agent-generated hypotheses were validated against human-generated ones. The trial-outcome association is also retrospective and correlational. That kind of head-to-head prospective comparison is what would settle how much autonomy these systems can be trusted with.
Two further pieces frame where this is heading. Krishnan, Collins and colleagues, in Science Translational Medicine, review AI for designing small-molecule and peptide antibiotics [5]. As a review, it adds no new data. Its value lies in flagging that data standards, reporting norms and regulatory pathways lag well behind the design capability. Song, Segal and Xing, in Nature Medicine, offer a perspective proposing an AI-driven digital organism [6]. This would be a set of linked foundation models simulating biology from molecules to whole individuals. It is a vision rather than a result. For clinicians, the practical reading of this whole cluster is that these tools may shape which drugs reach trials within years, but none has yet shown benefit in a patient.
Closer to the clinic again, two studies use machine learning to extract biomarkers from molecular data. He and colleagues, in Nature Medicine, applied a graph neural framework to more than six million nuclei from the prefrontal cortex of over a thousand brains [7]. Their aim was to link specific cell subpopulations to Alzheimer's phenotypes. They identified microglial subtypes tied to pathology and reactive astrocytes that may confer cognitive resilience. In cognitively impaired donors they found excitatory-inhibitory imbalance and mitochondrial dysfunction. They also found astrocyte signatures shared between Alzheimer's disease and depression. Findings replicated in external datasets, and the tool is openly available. But this is post-mortem, cross-sectional tissue work, so the associations generate targets rather than diagnostics.
Cascone and colleagues, in Nature, reported biomarker analyses from the CheckMate 77T trial of perioperative nivolumab in resectable non-small cell lung cancer [8]. Among patients on nivolumab whose circulating tumour DNA cleared before surgery, half achieved a pathologic complete response, compared with about one in eight on placebo. Patients who turned molecular-residual-disease positive during adjuvant treatment all relapsed. Tumours carrying KEAP1, STK11, CDKN2A or SMARCA4 alterations, often considered resistant to immunotherapy, appeared to benefit, with the risk of an event roughly halved. A machine-learning model ranked ctDNA clearance, non-N2 disease, complete response and squamous histology among the top predictors of event-free survival. The important caveat is that only about four in ten randomized patients had evaluable biomarkers, so these subgroup signals are hypothesis-generating. Notably, the AI here confirms markers clinicians already recognize rather than discovering new ones.
If you read only one paper from this period, make it Booth and colleagues on the ambient scribe [1]. It reopens the question of what scribes are actually for. They appear to save time on the note without moving closure rates or productivity, which shifts the debate from adoption to measurement.
Here is what this period adds up to. Ambient scribes now have large-scale real-world evidence of a modest, dose-dependent time saving, but the observational design leaves causation and downstream benefit unproven. Multimodal foundation models are performing well on benchmarks and offering visible reasoning, yet they still have no prospective clinical data. AI research agents are making bold claims about discovery and trial analysis, and the open question is how much autonomy they warrant, which only prospective validation will answer. And in oncology, ctDNA clearance continues to look like a strong early signal of benefit from perioperative immunotherapy, though in a biomarker subset too small to be definitive.
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.
Spot something worth flagging?
Get this every week in your podcast app — free.
New research episodes land in your feed automatically — listen on your commute.