Medical AI research, in brief
Skim recent peer-reviewed medical AI studies in plain language, from journals such as npj Digital Medicine, NEJM AI, The Lancet Digital Health, JAMA Network Open and JMIR. Each summary is generated from the publisher's abstract and approved by our team before it appears. Items stay for 60 days.
Get new studies and new tools by email, once a week.
JMIR cancer ·
Researchers developed a multilingual AI-driven system for cervical cancer awareness that combines a rule-based risk engine, a retrieval-augmented GPT-4o mini chatbot and a facility recommendation module. The RAG chatbot had retrieval precision@3 of 0.88, and clinician review found factual errors in 8% of its responses versus 28% for a non-RAG baseline. The authors state effectiveness in community settings remains to be evaluated.
Read the full article at JMIR cancerJAMA Network Open ·
In a 2-group randomized clinical trial of 400 pregnant and postpartum participants from minoritized groups, access to a generative AI maternal health chatbot was compared with an active control. The chatbot was not associated with differences in anxiety or depression scores at 12 months. The authors say secondary and exploratory findings should be interpreted cautiously.
Read the full article at JAMA Network Opennpj Digital Medicine ·
Researchers audited six open-weight large language models acting as judges of clinical text against expert judges on four German clinical datasets. Alignment with experts varied by criterion and corpus, all judges showed position and identity bias, and substituting LLM judges for experts missed 10.2% of clinical errors. The authors conclude LLM-as-a-judge is not a safe substitute for experts in German clinical text evaluation.
Read the full article at npj Digital MedicineBreast disease ·
This retrospective study compared the Lunit AI mammography system with single- and double-reader radiologists in a Singapore screening cohort of 1,982 women. Lunit AI reached an AUC-ROC of 0.874 versus 0.811 for single-reader and 0.863 for double-reader assessment, with higher specificity but lower sensitivity than single readers. The authors conclude it may be better suited as an adjunct than as a fully independent reader.
Read the full article at Breast diseaseRadiology: Artificial Intelligence ·
The abstract states that CT angiography-based AI showed high diagnostic performance for combined internal carotid artery and first-segment middle cerebral artery occlusions. Sensitivity was limited for distal second and third segment middle cerebral artery occlusions.
Read the full article at Radiology: Artificial IntelligenceRadiology: Artificial Intelligence ·
MultiModalNet is a deep learning model that combines longitudinal mammograms and clinical data to predict breast cancer recurrence and recurrence subtype. According to the abstract, it outperformed traditional methods.
Read the full article at Radiology: Artificial IntelligenceRadiology: Artificial Intelligence ·
DeepVEST is a deep learning model for vessel segmentation and removal on breast MRI. The abstract reports that its vessel-free maximum intensity projection images were rated by radiologists as effective, with minimal artifacts and substantial interreader agreement.
Read the full article at Radiology: Artificial Intelligencenpj Digital Medicine ·
Researchers developed and externally validated machine learning models to predict persistent fever 48-72 hours after antibiotics begin in hospitalized adults with cancer, using EHR data from two academic hospitals. The best model achieved an AUROC of 0.82 internally and 0.78 externally. The authors say use should be adjunctive only and that prospective studies are required.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
AnySleep is a deep neural network that scores sleep from any EEG or EOG data at adjustable temporal resolutions, trained and validated on over 20,000 overnight recordings from 28 datasets. The abstract reports state-of-the-art performance, matching or surpassing established baselines at 30-second epochs. On shorter timescales it improved prediction of obstructive sleep apnea, narcolepsy type 1 and insomnia, and the model is publicly available.
Read the full article at npj Digital MedicineRadiology: Artificial Intelligence ·
The abstract discusses automation bias and deskilling as risks of routine radiology AI use. It proposes a confidence-stratified triage model in which calibrated AI confidence routes cases to low-, intermediate- or high-confidence reading streams. It also proposes a research agenda to test whether these countermeasures reduce automation bias and skill erosion.
Read the full article at Radiology: Artificial Intelligencenpj Digital Medicine ·
This umbrella review included 50 reviews covering 387 primary studies on AI fracture detection. Pooled AI sensitivity was 0.897 and specificity 0.904, but certainty of evidence was very low to low for almost all outcomes and no review reported safety outcomes. The authors conclude real-world and comparative performance remain uncertain and call for prospective external validation and randomised trials.
Read the full article at npj Digital MedicineActa neuropsychiatrica ·
This study evaluated ChatGPT-4 responses to 25 frequently asked patient questions about rTMS, rated by three psychiatrists for quality, reliability and readability. Response quality was moderate to good (mean GQS 3.69), but the mean DISCERN score was 35.33 (poor) and the JAMA Benchmark score was 0 for all questions. The authors conclude ChatGPT may support patient education but should be used with caution and under clinician guidance.
Read the full article at Acta neuropsychiatricanpj Digital Medicine ·
This study developed and validated a fully automated deep learning pipeline that identifies C1–C3 vertebral levels of the upper cervical spinal cord on brain MRI, with a quality-control module. The best model (nnU-Net) reached Dice scores of 0.85–0.89, and external validation on 2,737 MRIs showed labeling accuracy above 92% in four of five datasets. The authors report sensitivity to longitudinal spinal cord atrophy in multiple sclerosis.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
The authors built a hybrid AI facial-recognition framework for Turner Syndrome detection, combining deep feature learning with geometric prior knowledge under data scarcity. It was developed retrospectively and validated in a prospective cohort, reaching 95.41% accuracy, 96.55% sensitivity and 95.24% specificity. The authors say the method offers a low-cost screening approach, particularly in resource-limited settings.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
This study compared automated metrics with blinded physician ratings for GPT-generated draft replies to patient portal messages, using 30 primary care physicians. Physicians preferred human-authored messages (4.24/5), and BERTScore correlated with physician quality ratings in the intervention group but not in the non-displayed group. The authors conclude that automated metrics have context-dependent validity and that clinical generative AI needs validation within its specific use context.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
The authors developed OUD-REY, a deep learning model that uses longitudinal EHR data from the All of Us program to predict one-year risk of opioid use disorder in patients with an opioid prescription. It achieved an AUROC of 91.2% on internal validation and 86.4% and 85.5% on temporally separated All of Us data and a VA dataset of about 6 million patients. Fairness auditing showed the model was generally fair across gender, race and ethnicity.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
The study tested whether large language models can rate depression from daily video diary transcripts in 108 participants. LLM ratings correlated with self-reported depression (r = .61 averaged across days) and with clinical-interview severity (r = .53). The authors report that LLM ratings predicted functioning and interview-rated depression beyond self-reports.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
The authors propose a multi-stage validation framework for LLM-based clinical information extraction and apply it to substance use disorder diagnoses across 919,783 clinical notes. Rule-based filtering and semantic grounding removed 14.59% of LLM-positive extractions, and the primary LLM reached an F1 score of 0.80 against judge evaluations. LLM-extracted diagnoses predicted later specialty care engagement more accurately than structured-data baselines.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
The authors present TrajOnco, a training-free multi-agent large language model framework that reasons over longitudinal EHRs for multi-cancer early detection. Evaluated on de-identified Truveta EHR data across 15 cancer types, it achieved zero-shot AUROCs of 0.64–0.80 and performed comparably to supervised machine learning in a lung cancer benchmark. Human evaluation supported the groundedness of its outputs.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
This benchmark study evaluated a large language model-based multi-agent framework for generating and refining optimization code for operating room scheduling. Across 10 LLMs and 35 scenarios, it had a higher feasible schedule rate than single-agent and traditional multi-agent baselines, and in 1050 synthetic cases plus three real OR datasets, it produced solutions close to expert-written mixed-integer programming models with improved OR use and less overtime in a hospital case study.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
This study developed RARE-PHENIX, an end-to-end AI framework that extracts phenotypes from clinical notes, maps them to Human Phenotype Ontology terms, and ranks rare disease-relevant terms. It was trained on 2,671 patients from 11 Undiagnosed Diseases Network sites and externally validated on 16,357 notes from 143 Vanderbilt patients, with findings showing high concordance with clinician curation and potential downstream disease-prioritization signal.
Read the full article at npj Digital MedicineJAMA Network Open ·
This cross-sectional study examined perceptions of using AI for medical patient portal messaging and mental health support, with opinions stratified by birth generation. Results are reported in the full article.
Read the full article at JAMA Network Opennpj Digital Medicine ·
This study used a Delphi process to develop expert consensus on the goals, use, and challenges of neuroimaging AI tools in memory clinics. Sixteen experts from clinical, academic, and industry settings completed three questionnaire rounds, reaching consensus on eight statements across purpose, quality, clinician and clinic interactions, and privacy and ethics.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
This evaluation examined whether electronic health record developers met federal transparency requirements for AI-enabled predictive decision support interventions. Across 35 developers, public reporting on risk analysis, mitigation, governance, and other required attributes was inconsistent and often incomplete, with wide variation in structure and detail, even though all developers were certified as meeting the requirements.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
The authors developed MOSAIC, an agentic AI system that codes clinical communication in patient–provider conversations using retrieval-augmented generation and a multi-agent workflow. Evaluated on 120 transcripts from rheumatology and OB/GYN against human coders, it achieved an overall weighted F1 score of 95.6%. The abstract describes this as showing potential for reliable clinical communication analysis.
Read the full article at npj Digital MedicineNature Medicine ·
This scoping review used a large-language-model-assisted pipeline to assess code sharing among 3,967 articles citing TRIPOD or TRIPOD+AI. Of these, 482 (12.2%) included code-sharing statements, and repository assessment showed substantial heterogeneity in reproducibility features. The authors say clearer expectations beyond code availability are needed, and the work informs the TRIPOD-Code guideline.
Read the full article at Nature Medicinenpj Digital Medicine ·
This PRISMA-guided systematic review evaluated AI deployments in radiation oncology in low- and middle-income countries; 18 studies met inclusion criteria. Most were task-level applications (50.0%) or clinician-supervised workflow integration (33.3%), none reached the highest integration level, and validation was mostly retrospective. The authors conclude that deployment remains at early stages and that safe scale-up needs advances in validation, workflow maturity, infrastructure and governance.
Read the full article at npj Digital MedicineJAMA Network Open ·
This randomized clinical trial tested a tailored chatbot plus a cancer predisposition clinic visit versus a clinic visit alone in 106 adolescents and young adults with hereditary cancer predisposition syndromes. The chatbot group showed greater improvement in accurate knowledge of cancer risk by age 30 years, while changes in screening knowledge and distress were not statistically significant; distress did not increase.
Read the full article at JAMA Network Opennpj Digital Medicine ·
This retrospective model development and validation study used EHR data from 373,389 adults with heart failure in the UK CPRD Aurum to create SIMPLE-HF, an 11-variable AI mortality risk score based on point-of-care factors such as age, body mass index, comorbidities, and medications. For 12-month all-cause mortality, SIMPLE-HF had better discrimination than an EHR-adapted MAGGIC score (C-index 0.801 vs 0.735), was well calibrated, and identified more true events at a 0.60 high-risk threshold.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
This prospective randomized single-blinded phase II trial compared LLM-assisted preoperative communication with standard communication in 268 patients with newly diagnosed prostate cancer scheduled for radical prostatectomy. Patients who received personalized LLM-generated responses before routine visits had lower anxiety, better emotion, illness perception, and satisfaction, while physicians had lower workload and shorter communication time.
Read the full article at npj Digital Medicinenpj Digital Medicine ·
This model development and benchmark evaluation study introduced LEME, an open-weight suite of ophthalmology large language models trained with instruction tuning and reinforcement learning. In zero-shot testing across five benchmarks, LEME outperformed seven baselines, including GPT-4o, and in three downstream tasks using deidentified patient data it received the highest clinician ratings for patient query responses, achieved the top F1 for visual acuity extraction, and generated assessment and treatment plans with ratings close to attending clinicians.
Read the full article at npj Digital MedicineJournal of Medical Internet Research ·
This systematic review and Bayesian diagnostic test accuracy network meta-analysis compared NCCT-based AI with expert and nonexpert readers for detecting anterior circulation large vessel occlusion using 10 retrospective validation studies. Unimodal imaging AI showed higher sensitivity than both reader groups with similar specificity, while clinically informed multimodal AI also had strong accuracy estimates but was supported only by indirect and sparse evidence.
Read the full article at Journal of Medical Internet ResearchNEJM AI ·
This independent external validation study evaluated Cytoreader-Fusion, an ensemble AI system for automated p16/Ki-67 dual-stain triage in HPV-based cervical screening, using 9006 cases across four blinded cohorts including a fully independent post-lock cohort of 4996 cases. The locked model showed stable performance across clinical settings and slide-preparation platforms, with sensitivity and specificity higher than ASC-US+ Pap cytology and comparable to or better than manual dual-stain interpretation; in the post-lock cohort, sensitivity for CIN 3+ was 88.3% and specificity was 54.4%.
Read the full article at NEJM AISummaries are written by Clinical Intelligence Foundry to help you find relevant research. They are not medical advice and may omit important details. Read the full source before relying on any finding. Clinical Intelligence Foundry is not affiliated with these publishers and does not endorse the products or studies mentioned.