AI in Medicine Digest – August 5, 2026
2026-08-05
- Nature Medicinea slide-level pathology foundation model trained on 2.3M whole-slide images and 14M question–answer pairs from 700,000 reports matches or exceeds the balanced accuracy of clinical-grade products for cancer detection in prostate, breast and breast lymph node — benchmark comparison, not a reader study or prospective deployment
- Nature Medicinein two experiments with 623 lay people and 153 primary care physicians, a fairness-constrained model improved accuracy and narrowed skin-tone disparities in both groups, but LLM explanations produced automation bias in lay users while physicians stayed resilient
- Radiologya breast-tomosynthesis foundation model pretrained on 25M sections from 487,975 volumes beat an ImageNet baseline on density classification (79% vs 73%) but showed no evidence of a difference on 5-year cancer risk (AUC 0.78 vs 0.76) or lesion detection (62% vs 67%)
- Naturea generative protein-design framework encodes any protein as a fixed-dimension probability distribution, shrinking proteins 10–25% and expanding EGF into variants with higher EGFR-binding affinity than wild type, validated in cells
- Nature Biomedical Engineeringan autonomous AI biologist trained on 24.4M publications nominated CHK2 as a Parkinson's target, and Chk2 inhibition rescued dopaminergic neuron loss and motor deficits in mice — preclinical
AI in Medicine Digest – July 30, 2026
2026-07-30
- Lancet Digital Healthan interpretable AI reanalysis of the only completed randomised trial in retroperitoneal sarcoma partitions the 266 patients into subgroups and finds two where preoperative radiotherapy meaningfully improves abdominal recurrence-free survival (3-year 79% vs 58%, HR 0.40, p=0.0016), while the STRASS- and STREXIT-defined subgroups do not hold up — derived and tested in the same cohort, so hypothesis-generating
- Science Advancesa DNA–carbon-nanotube sensor array read out by machine learning classified liver, lung and ovarian cancer from 253 serum samples at 89% sensitivity / 96% specificity for about $4 a test — a case–control feasibility study, not a screening trial
- npj Digital Medicine11 open-weight models rated 678 clinical-high-risk interview transcripts from 373 participants, best model 93% sensitivity but only 58% specificity, with 3% clinically relevant confabulation and site-level performance variation
- npj Digital Medicinein 163 volunteers receiving bad news, human-delivered consultations produced higher stress and salivary cortisol than AI-styled ones, but memory retrieval was worst in the AI-chatbot arm — a Wizard-of-Oz design, so it measures belief about AI, not any real model
AI in Medicine Digest – July 27, 2026
2026-07-27
- Nature Medicinean AI clinician-decision-support system predicts 17 inherited-retinal-disease genotypes from retinal images, and in a 295-patient randomized trial it lifted specialists' top-5 genetic accuracy from 67.3% to 88.5% (P<0.001) and improved downstream management decisions — a rare prospective RCT of diagnostic AI
- Nature Biomedical Engineeringa visual chain-of-thought pathology agent trained on expert whole-slide viewing behaviour outperformed state-of-the-art vision-language models on gastrointestinal lymph-node metastasis detection and held up on an independent external cohort — external but qualitative, no hard AUROC reported
- Nature Biomedical Engineeringan auditable chest-radiograph foundation model that decomposes every prediction into weighted radiological concepts, trained on 0.87M image-report pairs (239,391 patients) and externally validated on 4 physician-annotated datasets across the US, Europe and Asia — retrospective
- Lancet Digital Healtha wav2vec2-based speech model estimates Huntington's-disease severity from voice, matching clinician inter-rater agreement (ICC 0.87 vs 0.847) with external replication (ICC 0.67) — authors stress prospective validation is still needed
AI in Medicine Digest – July 22, 2026
2026-07-22
- npj Digital Medicinea histopathology-based survival system for pancreatic ductal adenocarcinoma, built across 5 cohorts (n=1,020), reaches a mean C-index of 0.761 with 6-month/2-year AUCs of 0.936/0.877 and stratifies patients by likely adjuvant-therapy benefit — retrospective
- npj Digital Medicinea multimodal (MRI + H&E + perioperative) deep survival model for hepatocellular carcinoma after curative resection, externally validated in a 6-center cohort (n=1,475, C-index 0.751), gives time-resolved 1–5 year risk and beats guideline staging, retrospectively
- npj Digital Medicinea contrastive chest-radiograph foundation model predicts 554 prevalent and 457 incident phenotypes across 3 cohorts (n>230,000) with explainable disease co-embeddings (R²≥0.85), a research tool pending prospective testing
- Nature Healthdynamic adversarial auditing of 15 health LLMs collapses 94% of previously-correct MedQA answers and elicits privacy leaks in 86% of scenarios, exposing a benchmark-vs-reliability gap
AI in Medicine Digest – July 16, 2026
2026-07-16
- Naturea Bayesian model jointly fitting longitudinal EHR diagnoses and polygenic risk across 3 biobanks (n>683,000, 348 diseases) outperforms the Pooled Cohort Equation, PREVENT and Gail at 1- and 10-year horizons
- Radiologyan eye-tracking reader study finds false-negative AI prompts collapse radiologist sensitivity from 71% to 39% and cut visual fixation on the cancers AI missed
- Nature Medicinea neuroimaging foundation model trained on 5.24M routine clinical MRI/CT volumes via 'health system learning' beats frontier models on report generation and reduces hallucinated findings, but the eval is retrospective/expert-preference
- npj Digital Medicinea transformer survival model for post-PCI ischemic/bleeding risk, externally validated in 19,173 patients (C-index 0.74-0.84), outperforms static DAPT scores
AI in Medicine Digest – July 10, 2026
2026-07-10
- npj Digital Medicinean AI algorithm deployed across 14 Polish health systems screened 1.3M patients and lifted the rare-blood-disorder detection rate to 10.9% vs 6.9% conventional, cutting diagnostic delays of up to 3.7 years
- npj Digital Medicinea prospective LLM voice assistant completed 88% of 1,431 real pre-procedural cardiac-cath calls over 6 months, with system errors falling to 2.6%
- Sciencean autonomous general-purpose biomedical AI agent mines tools across 25 domains and generalizes across gene prioritization, drug repurposing and rare-disease diagnosis, but validation is benchmark and case-study only
- Naturea self-supervised foundation model builds a universal embedding of 36M cells across 8 species with zero-shot transfer, a research tool not a clinical one
AI in Medicine Digest – July 4, 2026
2026-07-04
- Nature Medicinea locally deployable, case-grounded LLM agent for hematology tumor-board decisions reaches 81.8% concordance on 555 external cases and 82.8% in a prospective 1-month silent trial (64 consecutive cases), with hallucinations in just 2/664 cases (0.3%)
- Nature Medicinea pan-cancer foundation model predicts immune-checkpoint-inhibitor response from bulk tumor transcriptomes, beating 22 methods across 16 retrospective cohorts spanning 7 cancers and 6 ICIs (accuracy +8.5%, AUPRC +15.7%), with responders showing longer overall survival (HR 4.7) — but validation remains retrospective/computational
AI in Medicine Digest – June 30, 2026
2026-06-30
- Naturedeep learning on a Swedish ECG-to-death registry flags a 2.2% high-risk group (7.0% annual SCD), 86% missed by LVEF, externally validated in the US and Taiwan
- Nature Medicinepragmatic cluster-RCT in 16 Kenyan primary-care clinics (n=9,691) finds LLM assistance safe but no reduction in treatment failure
- Nature Medicinemultimodal ML across 1,714 immunotherapy patients identifies a histidine-linked metabolic signature of response, with modest external validation
- Naturemembership-inference attacks re-identify individual patients near-perfectly and hit underrepresented groups hardest
- Nature MedicineGPT-5 and Gemini ace health benchmarks but fail under small perturbations, exposing a benchmark-vs-readiness gap