9 articles · 3 topics
Critical care medicine2 articles
Score 0.56 · ✓ Verified
Renato Daltro-Oliveira et al. — Critical Care Medicine · October 19, 2026
Why it’s here: This article directly addresses ICU-acquired infections in critically ill patients, which aligns with the interest in ICU outcomes and management.
Study Design
This retrospective cohort study utilized prospectively collected data from 38,126 adult patients admitted to 50 ICUs in Brazil between October 2019 and December 2021. Patients included had an ICU stay exceeding 48 hours and no pre-existing infections within the first 48 hours. The primary outcome was the incidence of the first ICU-acquired infection (ICU-AI) during the ICU stay, analyzed using Fine-Gray subdistribution hazard ratios.
Key Results
In a multivariable competing-risk analysis, patients with hematological malignancies showed a subdistribution hazard ratio (sHR) of 1.64 (95% CI, 1.01–2.65) and those with other immunosuppressive conditions had an sHR of 1.61 (95% CI, 1.02–2.54) for ICU-AIs. Patients with solid malignancies did not have a significantly different risk (sHR, 0.88; 95% CI, 0.68–1.14) compared to non-immunosuppressed patients.
Why It Matters
The findings indicate that patients with hematological malignancies and other immunosuppressive conditions have a significantly higher risk of acquiring infections in the ICU. This highlights a critical vulnerability for these patient groups within the intensive care setting. The study did not find a significantly increased risk for patients with solid malignancies compared to non-immunosuppressed individuals.
Score 0.44 · ✓ Verified
Arunangshu Ghoshal et al. — BMC Health Services Research · October 12, 2026
Why it’s here: This study protocol addresses end-of-life and palliative care in the ICU, directly relevant to ICU outcomes and clinical management.
Study Design
This is a five-year, multicentre, multiphase, sequential mixed-methods implementation study designed to co-develop, implement, and evaluate a context-sensitive care package for people with advanced or terminal illnesses requiring ICU support in India. The study employed evidence synthesis, Delphi process, Nominal Group Technique, human-centred co-design, usability testing, piloting, process evaluation, and a quasi-experimental pre–post effectiveness design. It was informed by Medical Research Council guidance, the Consolidated Framework for Implementation Research, Normalization Process Theory, and the Implementation Research Logic Model.
Key Results
At the time of revision, Phases 1 and 2 of the study were ongoing, focusing on the development and piloting of the Multifaceted, Comprehensive, Appropriate Care Package (MCACP). Phases 3 and 4, which involve evaluating implementation across eight ICUs and assessing effectiveness with a planned sample of 150 patient episodes per ICU in pre- and post-implementation periods, had not yet commenced.
Why It Matters
The MATHRU study aims to evaluate the integration of an appropriate-care package across heterogeneous Indian ICUs, addressing inconsistent end-of-life and palliative care practices. The non-randomised design is acknowledged as susceptible to secular trends, regression to the mean, and residual confounding, necessitating interpretation alongside contextual data and implementation mechanisms.
Healthcare machine learning6 articles
Score 0.78 · ✓ Verified
Chun-Nan Hsu et al. — npj Digital Medicine · October 19, 2026
Why it’s here: Develops and validates a deep learning model for predicting opioid use disorder using EHR data, a direct application of predictive models in clinical care.
Study Design
This study developed OUD-REY, a deep learning model utilizing longitudinal electronic health record (EHR) data to predict opioid use disorder (OUD) risk at one year for patients with an opioid prescription. The model leveraged a pretrained language model to handle thousands of unseen, out-of-vocabulary clinical features within patient EHR histories.
Key Results
On internal validation, OUD-REY achieved an AUROC of 91.2% and AUPRC of 77.0%. External validation on a temporally separated All of Us dataset yielded an AUROC of 86.4% and AUPRC of 23.1%, while validation on a Veterans Affairs dataset of ~6 million patients resulted in an AUROC of 85.5% and AUPRC of 19.5%.
Why It Matters
The developed OUD-REY model demonstrates generalizability across different datasets and shows fairness for patients across various demographic groups. This model has the potential to improve the prevention and care of OUD.
Score 0.72 · ✓ Verified
An agentic AI system for multi-framework communication coding
Bohao Yang et al. — npj Digital Medicine · October 11, 2026
Why it’s here: Develops an AI system using LLMs for clinical communication coding, demonstrating real-world deployment and analysis in healthcare settings.
Study Design
This study developed MOSAIC, a Multi-framework Structured Agentic AI system for scalable Clinical Communication coding. MOSAIC combines dynamically few-shot prompting, retrieval-augmented generation (RAG), and a LangGraph-based multi-agent workflow for planning, annotation, and verification.
Key Results
MOSAIC was developed on 54 annotated transcripts and evaluated on 120 transcripts spanning rheumatology and obstetrics/gynecology (OB/GYN). On the evaluation set, MOSAIC achieved an overall weighted F1 score of 95.6%, with performance highest in rheumatology (F1 = 95.9%) and strongest for bias and patient behaviors categories (F1 = 96.4% and 96.1%, respectively).
Why It Matters
These results demonstrate the potential of multi-framework AI systems for reliable clinical communication analysis. MOSAIC addresses the labor-intensive, inconsistent, and difficult to scale nature of manual annotation for clinical communication.
● Preprint · Score 0.72 · ✓ Verified
Kagawa, E. et al. — medRxiv Preprint · cardiovascular medicine · October 19, 2026
Why it’s here: This article applies deep learning and AI to electrocardiogram interpretation for differentiating cardiac conditions, including explainability and human-AI interaction, aligning well with the researcher's interests.
Study Design
This study developed and externally validated an ensemble of three deep learning models for differentiating Takotsubo syndrome from acute anterior STEMI using 12-lead ECG images. A four-phase reader study was conducted with 70 healthcare providers to assess the impact of AI assistance and visual explanations (Grad-CAM) on diagnostic performance.
Key Results
The AI model achieved an external-validation accuracy of 0.914, and AI assistance significantly improved reader diagnostic correctness from 0.632 to 0.864 (OR, 3.71). Grad-CAM heatmaps provided an additional benefit (OR, 1.53), and less experienced readers showed greater improvement, with non-cardiologists outperforming cardiologists when using AI assistance.
Why It Matters
An ECG image-based deep learning ensemble effectively differentiated Takotsubo syndrome from anterior STEMI and enhanced reader accuracy, with visual explanations offering further benefit. The findings underscore the importance of effective human-AI interaction, particularly for less experienced clinicians, though AI-assisted readers remained less accurate than the model alone.
Score 0.66 · ✓ Verified
R. J. Waken et al. — Journal of the American Medical Informatics Association · October 18, 2026
Why it’s here: Investigates the real-world deployment and impact of an AI scribe in primary care, directly related to clinical care.
Study Design
This quasi-experimental study employed an interrupted time-series design to evaluate the longitudinal impact of ambient AI scribe use in primary care. The study included 220 clinicians across 36 clinics, analyzing 150 days prior and 150 days after their initial AI scribe access.
Key Results
At 150 days post-implementation, note writing time decreased by 16% (IRR 0.84), pajama time by 19% (0.81), and time outside of scheduled hours (TOSH) by 13% (0.87), while work relative value units (wRVU) increased by 2% (1.02). No changes were observed in note length at any time point.
Why It Matters
The study observed persistent changes in AI scribe use measures over time as clinicians integrated these tools into their workflows. While note length remained unchanged, reductions in documentation time and increases in wRVU were noted by day 150.
● Preprint · Score 0.64 · ⚠ Check
Bou Ali, M. N. et al. — medRxiv Preprint · endocrinology · October 19, 2026
Why it’s here: This study utilizes a fine-tuned large language model (LLM) and gradient-boosting for predicting glucose rise in type 2 diabetes based on meal descriptions, directly matching the researcher's interest in LLMs and predictive models in clinical care.
⚠ a section is not grounded in the abstract (The 'Study Design section' contains the sentence 'The estimated carbohydrate was then used with personalized gradient boosting to predict postprandial glucose rise in two independent cohorts: ShanghaiT2DM and CGMacros.' The abstract states that the model was transferred between datasets and that the full 'text [->] LLM [->] glucose' chain was evaluated in ShanghaiT2DM only. It does not explicitly state that the estimated carbohydrate was used with personalized gradient boosting in *both* cohorts, only that the prediction was compared in both. The 'Key Results section' contains the sentence 'Personalized gradient boosting achieved correlations of r = 0.63 in CGMacros and r = 0.49 in ShanghaiT2DM for predicting glucose rise, approximately doubling the correlation of a linear reference.' The abstract states that the correlation for CGMacros was r = 0.63 [0.55, 0.69] and for ShanghaiT2DM was 0.49 [0.42, 0.54]. The sentence in the 'Key Results section' omits the confidence intervals, which is a minor omission but makes it not *directly and unambiguously* supported. The 'Why It Matters section' contains the sentence 'While carbohydrate improved prediction in one cohort (CGMacros), its impact was not observed in the other (ShanghaiT2DM), highlighting dataset differences that require further investigation.' The abstract states 'Carbohydrate improved this prediction in CGMacros but not in ShanghaiT2DM; the two datasets differ in many respects, and the cause of this gap is not established.' This sentence is directly and unambiguously supported.)
Study Design
This study fine-tuned the Qwen3-4B language model using QLoRA on public data to estimate meal carbohydrate content from text descriptions. The estimated carbohydrate was then used with personalized gradient boosting to predict postprandial glucose rise in two independent cohorts: ShanghaiT2DM and CGMacros.
Key Results
The fine-tuned LLM achieved a mean absolute error of 12.2 g for real meals, nearly halving the error of a matched food composition table and significantly outperforming the unfine-tuned model. Personalized gradient boosting achieved correlations of r = 0.63 in CGMacros and r = 0.49 in ShanghaiT2DM for predicting glucose rise, approximately doubling the correlation of a linear reference.
Why It Matters
This approach demonstrates that a fine-tuned LLM can substantially improve carbohydrate estimation from meal text, aiding glycaemic control in type 2 diabetes. While carbohydrate improved prediction in one cohort (CGMacros), its impact was not observed in the other (ShanghaiT2DM), highlighting dataset differences that require further investigation.
✦ Something Different · Score 0.36 · ⓘ Reference
Extending synthetic medical education: generating OSCE cases with LLMs
Kaiz P. Esmail et al. — npj Digital Medicine · October 19, 2026
✦ A change of pace: This article explores the use of LLMs for generating synthetic medical education cases, offering a novel application of AI in training and assessment rather than direct clinical prediction or intervention.
Abstract unavailable — listed as a pointer; not summarized.
Hospital logistics and operations1 article
★ Flagship · Score 0.67 · ✓ Verified
Seokjun Youn, Yingchao Lan — Manufacturing & Service Operations Management · October 19, 2026
Why it’s here: Directly analyzes operational implications of diagnostic discrepancies and their impact on resource use and costs in hospitals.
Study Design
This study leverages a patient-level dataset of 5.7 million emergency-department-origin inpatient stays from Florida hospitals to examine the operational burden associated with diagnostic discrepancies. The primary method involves measuring this burden by total costs per hospitalization, adjusted for patient risk and high-dimensional fixed effects.
Key Results
Inpatient stays flagged with diagnostic discrepancies exhibit a 6.62% higher adjusted total cost compared to those without. This cost differential is smaller when physicians have higher experience with similar diagnoses (focal experience) and larger when they have higher experience with diverse diagnoses (related experience).
Why It Matters
The discrepancy flag is best utilized for retrospective monitoring, benchmarking, and case-review triage, rather than as a direct measure of preventable error or an automatic performance score. The observed patterns related to physician experience suggest potential areas for prospective evaluation in assignment, consultation, training, and decision-support practices.
Curated from OpenAlex, Crossref & medRxiv. Summaries are machine-generated and grounded in each article’s abstract or open-access text; ⚠ flags a summary that needs a second look, and ✦ marks a deliberate change-of-pace pick.
Leave a Reply