PubMed + Crossref
60 entries
Newest 14 Sep 2026
The paper feed
New research on AI in healthcare
Tracked from PubMed and Crossref across the journals that publish clinical AI. Every entry carries its citation and a link to the source.
Metadata only
Links out
Our own note
Recent work
What has published lately
AIMOCS lists bibliographic details and links out to the publisher. Abstracts and full text stay with the journals that hold their copyright.
Prospective validation of machine learning models predicting Gram negative bacteremia in ICU versus clinical scores
Multi-Agent collaboration as a complementary architecture for AI-generated medical examination items
A novel multiomics machine learning signature identifies rapid progression in clinically low risk prostate cancer
Data-driven identification of repurposable drugs for ALS
Causal machine learning for extracting insights from observational radiotherapy data
BenchECG and xECG: a benchmark and baseline for ECG foundation models
Institutional approaches to artificial intelligence policy and guidance in health informatics and information management education: emerging trends and inconsistencies
A scoping review of artificial intelligence-enabled wearables for medication adherence
Publisher Correction: A scoping review on the mental health harms of LLM-based chatbots
Clinical usability of an explainable AI decision support tool and evaluation of multimodal models in NSCLC
Development and validation of a machine learning-based predictive model for prognosis in cerebral hemorrhage patients after hyperbaric oxygen therapy
Safety, efficacy and acceptability of human-GenAI single-session exposure-based intervention for academic anxiety: randomized controlled trials
Deep learning combined habitat radiomics analysis of central lymph node metastasis in papillary thyroid carcinoma
Deep learning predicts gene rearrangements from histopathology in large B-cell lymphoma
Digitally enabled patient-centred survivorship care: from hype to implementation-aware impact
Evaluating AI-assisted detection of fetal intracranial malformations in prenatal ultrasound practice: a multicentre, self-crossover, randomised controlled trial in China
Vascular waveform analysis using Bayesian pulse deconvolution
Interpretable four-class machine learning prediction of initial I-131 therapy responses in differentiated thyroid cancer
Impact of LLM-supported patient education on patient perspectives and patient-reported outcomes: a mixed-methods systematic review
RDMA: cost effective agent-driven rare disease mining from electronic health records
A global framework for artificial intelligence education in medicine: international working group recommendations
Regulating artificial intelligence in health care: a tech-enabled, people-centred future
Impact of LLM assistance on physician decision-making: a multi-country randomized controlled trial
Our take
A randomised trial gave 249 physicians in Indonesia, Kenya and the Netherlands either standard conditions or access to GPT-4o while working through clinical vignettes, then compared performance. The control group had no internet or guideline access either, so this measures the model against nothing, not against a doctor's usual tools. LLM access raised scores everywhere, and the gain tracked the wealth of the setting inversely: Kenya rose 18 percentage points (95% CI 12.7–23.2, p<0.001), Indonesia 10.7 points (95% CI 5.7–15.7, p<0.001), and the Netherlands 7.2 points (95% CI 3.7–10.7, p<0.001). The pattern is the finding — a tool that narrows the gap between systems with different baseline resources, rather than lifting every setting by the same amount. The trial was funded on behalf of the Global Partnership on Artificial Intelligence and pre-registered with the AEA RCT Registry (npj Digital Medicine, 9 September 2026). It did not assess harms, and it tested written vignettes under controlled conditions, not decisions on real patients — so what it shows is potential, not proof of safe use at the bedside.
PREDICT-GBM: A multicenter platform advancing personalized glioblastoma radiotherapy planning
DystoniaDBSNet as a novel deep learning biomarker of predictive deep brain stimulation outcome in dystonia
A domain-specific large language model for evidence-based nursing: development and multi-method evaluation of EviNurse
AI-derived carotid plaque vulnerability from B-mode ultrasound for cerebrovascular risk stratification: a multicenter cohort study
Our take
Reading a carotid ultrasound for plaque risk is normally a judgment call, and this multicentre study tried to make that judgment reproducible. The team trained an AI model on 38,090 images from 6,618 people at high cardiovascular risk across four centres, checking plaque detection, five B-mode features, and an overall vulnerability score against expert readers. Detection scored an AUC of 0.95 internally and 0.96 on an external test set; vulnerability classification scored 0.90 and 0.88. The model outperformed six independent human readers assessing the same images. The more consequential test came in a separate group of 2,174 people followed for a median of 37 months, of whom 277 had a stroke or other cerebrovascular event: people the AI flagged as high risk had roughly double the event rate after adjusting for standard risk factors (hazard ratio 2.09, 95% CI 1.69–2.58), and adding the AI score to a clinical risk model raised its discrimination from a C-index of 0.769 to 0.795. That is a real but modest gain on top of what clinicians already assess (npj Digital Medicine, 8 September 2026).
Intervention fidelity and competency translation of serious games in pharmacy education: a scoping umbrella review
Our take
This is a review of reviews, not a new trial: the authors searched three databases through the end of 2025 for systematic reviews of game-based teaching in pharmacy education, screened 914 unique records, and kept ten — eight found by database search, two by tracing citations backward. Their interest was whether these games change what pharmacists actually do at work and for patients, not just what students score on a quiz right after playing. The answer, on the evidence assembled here, is that engagement and short-term knowledge are what gets measured and reported; workplace behaviour and patient outcomes are rarely tracked. Seven of the ten included reviews pooled digital and physical game formats together, which means their findings cannot be pinned specifically on the digital versions. No effect size for behaviour change is reported, because the underlying reviews mostly did not have one to report (npj Digital Medicine, 8 September 2026). The honest reading is a gap in the evidence, not a result either way.
International qualitative case studies of system-level approaches to promote the development, adoption, and implementation of artificial intelligence in healthcare
Our take
A qualitative study, not a trial, so there is no effect size to report here — the finding is a pattern across four health systems that have tried to move AI in healthcare past pilot projects: Catalonia, Norway, Singapore and Queensland. The researchers drew on 60 documents, 34 interviews and 5 focus groups with 50 policymakers, strategic decision-makers and lead clinicians across the four sites. Across all four, how far AI scaled depended less on the technology than on each system's existing digital strategy, funding structure and legacy infrastructure, and on whether governance, procurement and post-deployment monitoring were built in as standing capabilities rather than one-off project tasks. The authors' main recommendation is that health-system leaders stop framing scaling as either a top-down mandate or a bottom-up pilot, and instead treat it as ongoing coordination across many stakeholders at once (JAMIA, 8 September 2026). It is a lessons-learned paper: useful for planning, but it makes no claim about clinical outcomes.
Health artificial intelligence is here, but are we measuring what matters?
Our take
A perspective piece, not a study — it reports no data of its own, so there is no sample size or effect size to cite here. Its argument: health AI evaluation still leans on technical metrics like accuracy and validation studies, when the question that actually matters is whether a deployed system changes a clinical or operational outcome, and the gap between how systems are evaluated and how their value shows up in practice is where much of the current disappointment with health AI comes from. The authors propose shifting toward outcome-based, systems-level evaluation: standardised outcome frameworks, dedicated evaluation infrastructure, governance that includes more than one stakeholder group, and incentives that reward measuring real impact rather than passing a benchmark (JAMIA, 8 September 2026). Read it as an argument for a different evaluation standard, not as evidence that any particular AI tool works or does not.
Evaluating retrieval-augmented generation versus long-context input for clinical reasoning over electronic health records
Our take
This study asks a practical question for anyone building a clinical AI tool on hospital notes: is it worth building a retrieval system, or can you just feed the model more text? The authors tested three large language models — GPT-5.4-mini, Mistral Medium 3 and DeepSeek V3.1 — on real inpatient notes from a US academic health system, across three tasks: pulling out imaging procedures, building a timeline of antibiotic use, and naming the key diagnoses of a hospital stay. For imaging extraction, retrieval-augmented generation (RAG) beat feeding in the most recent notes directly, and even beat giving the model the full long context, by 0.17 to 9.83 F1 points across the three models, using under 8,000 tokens. Antibiotic timelines showed a similar pattern, with under 8,000 retrieved tokens matching long-context performance (a Jaccard-score difference of –3.26 to +3.24). Diagnosis generation was the exception: no method or model moved the needle much, which the authors attribute to gaps in what the notes actually document, not a modelling failure (JAMIA, 8 September 2026). The takeaway is task-dependent, not a blanket case for retrieval.
An explainable detection framework for health insurance fraud via temporal capture and confidence assurance
Our take
A fraud-detection model built for health-insurance claims, not for patient care directly. The authors trained a model called TADCA — a memory network reading the sequence and timing of a provider's claims, with an attention layer marking which claims looked most suspicious — and tested it against nine other machine-learning and deep-learning baselines. Under a fixed audit budget of reviewing the top 10% of flagged providers, TADCA scored the highest AUROC of the group at 0.907, with a Precision@10% of 0.624 and a Recall@10% of 0.648 — meaning roughly six in ten providers flagged in that top slice were confirmed fraudulent, and about the same share of true fraud cases were caught within it. The paper also adds a calibrated uncertainty layer around each prediction, with coverage reaching 0.992 in one configuration, meant to help auditors decide which flags to trust (npj Digital Medicine, 8 September 2026). It is a detection tool for payers and auditors, not a clinical one, and the paper does not report how it performs outside the claims data it was trained and tested on.
Identifying potential nonpulmonary vein triggers in persistent atrial fibrillation using digital twins and deep learning
Our take
Persistent atrial fibrillation often continues after doctors isolate the pulmonary veins, and the other trigger sites elsewhere in the atrium are hard to find without invasive testing or a patient-specific fibrosis model that is slow to build. This study trained a deep-learning model, InduceNet, on results from personalised digital heart simulations to flag those non-pulmonary-vein trigger sites without needing the fibrosis-modelling step. A broader set of simulated pacing sites turned up potential trigger sites in 85% of the digital hearts tested, against 54% for the sites a standard catheter procedure would normally test — and InduceNet predicted which sites those were with 91% sensitivity. The work stops at simulation and prediction: it does not report ablation outcomes in real patients guided by the model, so it establishes a faster route to a candidate map of triggers, not yet a tested treatment strategy (npj Digital Medicine, 8 September 2026). Whether finding more candidate sites changes ablation success is still an open question.
How Repeal of the NTAP Alternative Pathway will Impact U.S. Clinical AI Innovation
Multimodal artificial intelligence integrating peritumoral MRI and clinical data stratifies meningioma recurrence risk after radiotherapy
Our take
A retrospective study of 250 people with meningioma treated with adjuvant radiotherapy across three neurosurgical centres, testing whether combining MRI of the tissue just outside the tumour with clinical data can predict who will have the tumour come back. The model examined three MRI sequences across three margins around the tumour site and found the 1-centimetre margin carried the most transferable signal across centres. Radiomics — hand-engineered image features — scored well on the training data but did not generalise well externally; features learned directly by a deep-learning network held up better. Combining the deep-learning imaging features with clinical and pathology data in one model gave the best result, a mean C-index of 0.856 across the three-centre cohorts, meaning it separated higher- from lower-risk patients reasonably well, though not perfectly. It also showed favourable calibration and grouped patients into distinct risk tiers (npj Digital Medicine, 7 September 2026). Retrospective and modest in size — a candidate tool for planning follow-up, not yet validated prospectively.
Unsupervised machine learning for placental disease using cell spatial organization
Senescence-aware filtering facilitates robust prediction of cancer immunotherapy outcomes
Our take
Models that try to predict who will respond to immune-checkpoint cancer therapy from a tumour's molecular profile often fail to generalise, because tumours with similar-looking molecular patterns can behave very differently — a form of biological noise the authors trace partly to cellular senescence. Their fix was to identify and exclude tumours showing senescence-associated features before training the predictive model, using a network-based analysis of immune-checkpoint and senescence-pathway activity. Tested across melanoma, gastric, bladder and lung cancer cohorts, the filtered approach improved accuracy, AUROC, precision and specificity compared with training on the full unfiltered set, and held up in an external melanoma validation cohort — though the abstract does not report the specific values for any of those metrics, so the size of the improvement cannot be quoted here. The excluded tumours were enriched for senescence markers, which the authors offer as evidence that the filtering step is removing a real biological confounder rather than noise (npj Digital Medicine, 5 September 2026). What is missing is a head-to-head comparison against simpler filtering rules, so it is not yet clear the senescence framing is doing the work.
From feasibility to neuroanatomic validity of remote cognitive smartphone assessments in early Alzheimer’s disease
Less can be better: decomposing clinical data modalities in large language model-based healthcare applications
Our take
Adding more data types to a clinical AI model sounds like it should always help; this study, built on 22,254 hospital admissions from the public MIMIC-IV, MIMIC-IV-Note and MIMIC-CXR-JPG datasets, tested when that is actually true. For predicting in-hospital mortality and length of stay, structured EHR data alone matched or beat models that also used radiology reports or chest X-rays — the best mortality AUROC was 0.849 and the best length-of-stay AUROC was 0.868, both from structured data, with little gained by adding the other modalities. For two decision-support tasks — predicting the discharge diagnosis and predicting medication use — combining all three data types did help substantially, reaching F1 scores of 0.589 and 0.405, a 21.4% and 18.4% improvement over the best single-modality model respectively. Radiology reports consistently added more than raw chest X-ray images did. The authors' conclusion is not "more data is better" but "match the data you add to the task" (JAMIA, 5 September 2026).
Exploring generalizability and explainability of LLMs in classifying clinically rated suicidal ideation using heterogeneous data
Our take
An AI classifier trained to spot clinician-rated suicidal ideation in clinical transcripts performed unevenly across patient groups — the kind of gap that matters most in exactly the setting where missing a case is costly. The authors found that a standard, topic-general classifier was substantially worse at detecting suicidal ideation in some subgroups than others; for people with alexithymia, a condition marked by difficulty naming or describing feelings, the model's odds of missing a genuinely suicidal case were 2.39 times higher than for people without it (p=0.002). Splitting the classification task into narrower, topic-specific classifiers narrowed that gap: the accuracy difference between subgroups, measured as a change in AUC, fell from 0.11 to 0.01 in one comparison and to 0.05 in another. The result argues that a single general-purpose model can quietly underperform for people who describe distress differently, and that decomposing the task by topic is one way to catch that before deployment — a proposed fix, not yet a validated screening tool (npj Digital Medicine, 5 September 2026).
Digital pathology, image analysis, and artificial intelligence in liver disease
A voice-biomarker foundation model for ALS monitoring and Parkinson’s screening
Platforms for artificial intelligence-enabled infectious disease surveillance
Early Radiographic Changes Predict Radiographic Stability in Distal Radius Fractures: Development and Validation of an Explainable Predictive Model
Organoid Intelligent Morphomics: Decoding the organoid morphome through artificial intelligence from phenotypic quantification to mechanistic insight
Differences in tone of AI and care team responses to patient messages by patient demographics
Explicit chain of diagnosis for CMR semantic interpretation via large reasoning model
Interpretable machine learning for individualized survival prediction in node-positive, non-metastatic prostate cancer: a population-based study
Author Correction: Effect of a digital health behaviour change support system on cardiovascular disease risk in a randomized weight loss trial
Status, challenges, and future directions of machine learning in the management of epilepsy: a systematic review and meta-analysis
The illusion of clinical reasoning: a benchmark reveals the pervasive gap in vision-language models for clinical competency
A five-phase evaluation framework for diagnostic and predictive medical artificial intelligence
Effect of a computer-aided quality feedback system on colonoscopists' adenoma detection rate in Denmark: a multicentre, stepped-wedge, cluster-randomised, controlled trial
AI models for medication adherence prediction: closing the gap to clinical readiness
Large Language Models Versus Multidisciplinary Tumor Board Decisions in Thyroid Cancer
Machine learning-ready genomic biomarkers: ATF3 polymorphisms predict postoperative analgesic demand through AI-compatible phenotyping
Development and benchmark validation of PubChat for PubMed-grounded multilingual biomedical literature retrieval
Deep learning from retinal images for prediction of biomarker-informed obesity subtypes associated with cardiometabolic risk
Development and external validation of a contrastive learning foundation model for ECG-based prediction of cardiovascular diseases and outcomes
The weekly read
One of these papers, read closely, every week.
The feed lists what published. The weekly email takes one paper apart: what it measured, how well, and whether it changes anything.
Join the network
One email address