Menu
Medical AI: What the Research Shows
Published studies from leading research teams on how accurate AI is in medical diagnosis, with a link to every source so you can read them yourself.
AI 4x More Accurate Than Physicians on Complex Cases
In Microsoft's MAI-DxO study, AI diagnosed 85.5% of 304 complex NEJM cases correctly, compared to 20% for physicians working without colleagues or references — at 20% lower cost.
“…this is a really big jump.” — Dr. Eric Topol, Scripps Research Translational Institute, speaking to TIME
What Published Research Shows About Medical AI
These aren't our claims — they're findings from studies by leading institutions worldwide, linked below.
vs 44% for Human Readers
Breast cancers detected across 115,973 mammograms
Google & Imperial, Nature Cancer 2026 →vs 64% for Cardiologists
Structural heart disease detected from routine ECGs
Columbia University study →vs 67% for Physicians
AI recommendations rated optimal in 461 visits
Cedars-Sinai, Annals of Internal Medicine →Fewer Diagnostic Errors
Clinicians using an AI copilot in real clinics
OpenAI + Penda Health (39,849 visits) →2026 Breakthrough Studies
The latest peer-reviewed research from early 2026, published in top-tier journals including Nature, Nature Medicine, and Nature Cancer.
Nature Medicine — LLMs as Medical Assistants for the Public
Randomized Preregistered Study
February 2026 • Peer-Reviewed
Tested alone, the AI models identified the right condition in 94.9% of cases. But members of the public using the same chatbots found it in fewer than 34.5% of cases, no better than people using any other source. How people use AI matters as much as how capable the model is.
Key Findings:
- • 1,298 participants, randomized and preregistered
- • AI alone: correct condition in 94.9% of cases
- • People using general chatbots: under 34.5%, no better than control
- • Published in Nature Medicine Vol. 32, pp. 609–615
Google — Mammography AI Screening
Multicenter Retrospective & Prospective Studies
March 2026 • Nature Cancer
Across 115,973 mammograms from five NHS screening services, Google's AI detected more cancers than the first human reader (54.1% vs 43.7% sensitivity) with comparable specificity, then ran live at 12 sites.
Key Results:
- • 115,973 mammograms, 39-month follow-up
- • Sensitivity 54.1% vs 43.7% for the first reader
- • Specificity non-inferior (94.3% vs 95.2%)
- • Prospective feasibility deployment at 12 sites (9,266 cases)
Stanford/Google — Merlin CT Foundation Model
Vision–Language Model for CT Imaging
2026 • Nature
Merlin learns from 3D CT scans together with radiology reports and health records. Evaluated on about 50,000 CT scans from four sites, it outperformed existing radiology foundation models by about 30%.
Key Highlights:
- • 3D vision–language model for computed tomography (CT)
- • Trained on 15,331 CT scans with reports and diagnosis codes
- • Evaluated on ~50,000 abdominal and chest CTs across 4 sites
- • ~30% higher F1 and AUROC than prior radiology models
DeepRare — Rare Disease Diagnosis
Multi-Agent LLM System
2026 • Nature
Across 2,919 rare diseases in 14 specialties, the DeepRare multi-agent system ranked the correct diagnosis first in 57% of cases, well ahead of the next-best method, and rare-disease physicians agreed with 95.4% of its reasoning chains.
Key Results:
- • 2,919 rare diseases across 14 specialties
- • 57.2% top-1 accuracy, 23.8% ahead of the next-best method
- • 95.4% agreement from 10 rare-disease physicians reviewing 180 cases
- • Traceable reasoning linked to medical evidence
2025 Breakthrough Studies
The most significant AI medical research published in 2025, demonstrating major advances in diagnostic accuracy.
Microsoft Research — MAI-DxO
AI Diagnostic Reasoning Agent
July 2025 • arXiv Preprint
Microsoft's AI diagnosed 85.5% of 304 complex NEJM cases correctly, compared to 20% for physicians working without colleagues or references — about 4x more accurate, at 20% lower cost.
Key Results:
- • 85.5% AI accuracy vs 20% for 21 physicians
- • About 4x more accurate on these cases
- • 20% lower cost to reach diagnosis
- • 304 New England Journal of Medicine case records
- • Dr. Eric Topol to TIME: “…this is a really big jump.”
Stanford Medicine — Randomized Trial
Nature Medicine Study
February 2025 • Peer-Reviewed
Physicians given GPT-4 scored 6.5% higher on clinical management reasoning than those using conventional resources, and GPT-4 on its own scored about as well as the physicians who used it.
Key Findings:
- • 92 practicing physicians, randomized
- • +6.5% management reasoning score with GPT-4
- • GPT-4 alone performed on par with physicians using it
- • Published in Nature Medicine (February 2025)
Real-World Clinical Implementations
OpenAI + Penda Health
AI Consult - Kenya Implementation
39,849 Real Clinical Visits
Clinicians using AI Consult made 16% fewer diagnostic errors and 13% fewer treatment errors, with the AI working as a co-pilot rather than a replacement.
Key Results:
- • 16% relative reduction in diagnostic errors
- • 13% relative reduction in treatment errors
- • 39,849 visits across 15 clinics, 106 clinicians
- • AI as clinician co-pilot, not replacement
Cedars-Sinai Medical Center
CS Connect AI Platform
Los Angeles, CA - 42,000+ Patients
In 461 virtual urgent-care visits, an expert panel rated AI recommendations optimal 77% of the time, compared with 67% for the treating physicians' decisions.
Study Details:
- • 77.1% of AI recommendations rated optimal vs 67.1% for physicians
- • Rated potentially harmful: 2.8% for AI vs 4.6% for physicians
- • Published in Annals of Internal Medicine (2025)
- • K Health platform; 42,000+ patients served
Global Research & Academic Studies
Columbia University
EchoNext AI Heart Disease Detection
NewYork-Presbyterian - 85,000 ECGs Study
The EchoNext AI detected structural heart disease from routine ECGs with 77% accuracy, compared with 64% for 13 cardiologists.
Study Highlights:
- • 77% AI accuracy vs 64% cardiologist accuracy
- • 85,000 ECGs analyzed in the study
- • Identified previously undiagnosed cases
- • Routine ECG transformed into screening tool
Reuters Health Rounds
EchoNext AI Heart Screening
Columbia University Study Coverage
Reuters' coverage of the same study: 77% accuracy for the AI versus 64% for cardiologists, with thousands of patients flagged whom traditional screening had missed.
Key Findings:
- • 77% AI accuracy vs 64% cardiologist accuracy
- • Thousands of patients flagged by AI
- • Missed cases identified by traditional screening
- • Expands usefulness of common heart tests
Academic Meta-Analyses & Peer-Reviewed Studies
npj Digital Medicine
Meta-Analysis of 83 Studies
2018-2024 Research Review • Nature Portfolio
Across 83 studies, generative AI averaged 52.1% diagnostic accuracy, with no significant overall difference from physicians, though it trailed expert physicians by about 16 percentage points.
Study Scope:
- • 83 studies analyzed (2018-2024)
- • Comprehensive AI vs physician comparison
- • Published in npj Digital Medicine (Nature Portfolio)
- • Trailing expert-level physicians by ~16 points
August AI (arXiv)
Health AI Chatbot Benchmark
400 Clinical Vignettes
The health AI “August” named the correct diagnosis first in 81.8% of 400 clinical vignettes across 14 specialties, asking 47% fewer questions than conventional symptom checkers.
Research Details:
- • 400 validated clinical vignettes, 14 specialties
- • 81.8% top-1 diagnostic accuracy (327/400)
- • 47% fewer questions than symptom checkers (16 vs 29)
- • Evaluated by the product's own developer; not peer-reviewed
Medical Imaging AI
Industry Roundups
Tumor & Disease Detection
Industry roundups report 94–95% accuracy for AI in identifying tumors, lung nodules and retinal disorders in published studies.
Research Coverage:
- • 94-95% accuracy reported in tumor detection
- • Lung nodule identification systems
- • Retinal disorder screening
- • Summaries by GlobalRPH and RamSoft (not peer-reviewed)
Leading Medical AI Researchers
Summaries of their published views, with links to the original articles. They are not affiliated with LabLumen and have not reviewed it.
Eric Topol, MD
Cardiologist & Digital Medicine Researcher
Scripps Research Translational Institute
Argues that AI, especially deep learning, can make diagnosis more precise and improve patient care.
Key Contributions:
- • Author of "Deep Medicine"
- • Leading voice on responsible medical AI
- • Advocates for AI humanizing healthcare
- • Director at Scripps Research Institute
Source Articles:
Maryellen L. Giger, PhD
Co-founder of Quantitative Insights
University of Chicago
Her work shows AI can match or exceed human-level accuracy in breast cancer risk assessment and treatment response analysis.
Key Contributions:
- • Developed QuantX (FDA-cleared ML system)
- • Pioneer in breast cancer diagnostics
- • Computational imaging research leader
- • First FDA-cleared ML system for cancer
Source Article:
UChicago Medicine FeatureAnant Madabhushi, PhD
Director of AI Research
Emory University & Georgia Tech
His research shows measurable accuracy gains from AI in image-based diagnosis, particularly in oncology.
Key Contributions:
- • Computational pathology leader
- • AI in oncology and cardiovascular disease
- • Image-based diagnosis and prognosis
- • Director at Emory & Georgia Tech
Source Articles:
Bringing Medical AI to Veterinary Care
VetLumen runs the same kind of AI models with veterinary-specific instructions, built-in lab reference ranges for dogs and cats, and your pet's species, breed, age and weight. That matters because normal values differ by species: a cat's resting heart rate (120-140 bpm) is very different from a human's (60-100 bpm), and normal kidney values differ between dogs and cats.
Dogs
Built-in canine lab reference ranges, plus breed context such as hip dysplasia in German Shepherds or heart disease in Cavaliers
Cats
Built-in feline lab reference ranges, with attention to kidney markers, hyperthyroidism and diabetes
Horses & Livestock
Species-aware analysis for horses and farm animals, drawing on the AI's general veterinary knowledge (no built-in lab reference ranges for these species yet)
Veterinary AI Research & Industry Adoption
Deep Learning in Veterinary Diagnostics
Frontiers in Veterinary Science comprehensive review of AI applications across veterinary medicine, including diagnostic imaging, pathology, and clinical decision support.
Read StudyAI Poised to Transform Veterinary Care
American Veterinary Medical Association reports on AI adoption in veterinary practices, including diagnostic imaging, medical records, and early disease detection tools.
Read ArticleRenalTech: AI for Feline Kidney Disease
Mars Petcare's AI tool trained on hundreds of thousands of cat medical records can predict chronic kidney disease up to 2 years before clinical diagnosis.
Read AnnouncementVeterinary AI Diagnostics Market
The AI in animal health market is projected to reach $7.5 billion by 2033, growing at 20.1% CAGR as practices adopt AI-powered diagnostic tools.
View Market ReportFirst AI-Designed Personalized mRNA Cancer Vaccine for a Dog
Australian tech entrepreneur Paul Conyngham used ChatGPT and AlphaFold to design a personalized mRNA cancer vaccine for his rescue dog Rosie, who had terminal mast cell cancer. The tumor shrank by 75% within one month. Scientists at UNSW and the University of Queensland produced and administered the vaccine — the first personalized cancer vaccine ever designed for a dog. Highlighted by Sam Altman and covered by Fortune, The Scientist, and more.
Put Medical AI to Work on Your Own Results
Join 3,700+ people who have run 6,400+ analyses with LabLumen. It uses the same kind of large language models studied above to help you understand your results before you talk to your doctor.