JAMA published a large-scale usage analysis of a clinical AI tool this week. Across 952,008 verified US clinicians and 271 million queries (January 2024 to June 2026), OpenEvidence grew from 355 to 703,265 monthly active users, queries per active user rose from 18.0 to 39.5, and 92.5% of June 2026 users were returning rather than new. Adoption is habitual, not experimental.
The content of those queries is the more consequential finding. Physicians used the platform mainly for clinical decision-making (51.9%), led by treatment and management (33.1%); diagnosis accounted for 13.9%. Most published clinical AI evaluation still benchmarks diagnostic reasoning and knowledge recall. Use skews toward early-career physicians (use ratio 2.54) and residents and fellows (1.84), and nurse practitioners and physician assistants query more intensively than physicians (44.5 vs 37.1 queries per active user-month). Procedural and diagnostic specialties barely register: orthopedics and radiology both sit at a use ratio of 0.24. Between 35% (oncology and hematology) and 88% (emergency medicine) of specialist queries fall outside the clinician's own specialty.
The timing is commercially relevant. Becker's Hospital Review reported on 28 September that OpenEvidence raised $250 million at a $15 billion valuation; this paper is now the most detailed public evidence of the engagement behind that figure. For health-system buyers, the implication, which the authors also draw, is that evaluation and governance should be anchored on treatment and medication questions rather than diagnostic benchmarks.
Conflict of interest and limits. Three of the seven authors are OpenEvidence personnel (a salaried employee, a full-time employee and the chief medical officer), and the company supplied the data and contributed to the analysis. The study covers a single platform, captures only queries entered in its main interface, and cannot separate point-of-care decisions from background reading. Queries were classified by an LLM, with 74.5% agreement with clinician labels at the subcategory level (82.3% at the five higher-order categories). The paper measures use. It does not measure answer accuracy or patient outcomes.
Pre-Print Intelligence (arXiv)
Verification Turns Capability Gaps into Routing Decisions: Small Language Models for Clinical Documentation
Brief: The paper proposes a verification‑driven routing architecture that lets a small 4B‑parameter language model (with QLoRA) generate clinical notes from doctor‑patient dialogue, escalating to a frontier model only when verifiers flag deficits. On a test set of 250 encounters, the small model achieved 0.984 claim‑level factual entailment (vs. 0.992 for the frontier) and comparable verified‑rate acceptance (99.6% vs. 98.4%) while avoiding frontier generation calls in 62% of cases. The approach runs entirely on‑device and isolates errors to inter‑section structural fidelity, which becomes an explicit routing decision rather than a silent quality gap.
Methodological Integrity: Evaluation relies on a single, relatively small test set (250 encounters) and a fixed verification stack, limiting evidence of generalizability across diverse specialties, patient demographics, or note types. The study does not report external validation, potential bias in the verification policies, or ablation of verifier performance, raising concerns about overfitting to the specific corpus and verifier design.
Strategic Implication: By reducing reliance on large‑scale cloud models for routine documentation, the method could lower compute costs and latency while maintaining safety‑critical factual accuracy, aligning with ambient, proactive AI principles in orthopedic and MSK workflows. If integrated into EHRs or dictation systems, it may enable scalable, on‑device clinical note generation that augments clinician throughput while retaining near-frontier factual entailment (0.984 vs 0.992).
Executive Summary: The verification‑augmented small language model achieves near‑frontier factual correctness in clinical note generation and cuts frontier model usage by over half through explicit routing. Its on‑device operation and error‑aware design address key safety and efficiency barriers in automated documentation.
Innovation: 8/10 | Applicability: 9/10 | Commercial Viability: 7/10
FOCUS: A Subspecialty Foundation Model for Intraoperative Central Nervous System Pathology
Brief: FOCUS is a foundation model pretrained on 27,930 intraoperative central nervous system frozen-section images, enabling accurate hierarchical diagnosis and biomarker prediction. It achieves mean AUCs of 0.969 on diagnostic tasks and 0.877 on biomarker tasks, and improves junior neuropathologists' lymphoma detection sensitivity from 81.6% to 94.7% in reader studies.
Methodological Integrity: The validation relies on internal cohorts and external temporal sets, but lacks independent multi‑institutional testing and detailed reporting of preprocessing steps that could introduce leakage. Reader studies involve only six neuropathologists, limiting generalizability of the observed performance gains.
Strategic Implication: If integrated into pathology workflows, FOCUS could reduce intraoperative diagnostic uncertainty and support neurosurgical decision‑making, particularly for lymphoma detection. However, its impact remains confined to specialized neuro‑oncology centers unless broader anatomic pathology adoption is achieved.
Executive Summary: FOCUS demonstrates strong diagnostic performance on CNS frozen‑section data and improves reader accuracy in a controlled reader study. The model requires further validation and workflow integration to assess real‑world utility.
Innovation: 9/10 | Applicability: 7/10 | Commercial Viability: 6/10
Dense Endoluminal Mapping and Localisation from Monocular Bronchoscopy
Brief: The paper presents a monocular SLAM system for bronchoscopy that estimates depth using a fine-tuned Depth Anything 3 backbone and matches features with DINOv3, integrating projective ICP constraints into a factor graph to localize the scope and reconstruct the airway surface from RGB video alone. It demonstrates full trajectory completion on six synthetic patient-derived airways and two ex vivo specimens, achieving 26.1 mm mean absolute trajectory error and 77.5% mean surface F-score.
Methodological Integrity: Validation relies heavily on synthetic airway renderings for depth network training and limited ex vivo specimens, raising concerns about domain shift to diseased or in vivo human airways. The system's scale is derived from a single depth estimator, and no real-time performance or robustness to respiratory motion is reported.
Strategic Implication: If accuracy can be improved to clinically relevant levels (<10 mm error), the approach could reduce reliance on electromagnetic trackers and preoperative CT, lowering setup cost and complexity for bronchoscopic navigation. However, current error margins likely exceed the precision needed for targeted biopsy, limiting immediate clinical utility without further refinement.
Executive Summary: The method achieves complete airway reconstruction and localization using only monocular bronchoscopy video, outperforming several monocular SLAM baselines on synthetic and ex vivo data. Its translational impact hinges on reducing tracking error and validating in vivo.
Innovation: 7/10 | Applicability: 6/10 | Commercial Viability: 7/10
Tool-Augmented Large Language Models for Reliable Medical Calculation from Patient Notes
Brief: The paper introduces a tool-augmented LLM framework that extracts calculator specifications from medical notes, builds executable tools, and performs deterministic calculations to improve reliability. It separates language understanding from rule execution, achieving 86.73% final-answer accuracy on the MedCalc-Bench dataset, outperforming baseline LLM approaches by significant margins.
Methodological Integrity: Evaluation relies solely on the MedCalc-Bench benchmark, which may not reflect the full variability of real-world clinical notes and introduces potential overfitting to specific calculator formats. No external validation or prospective clinical study is reported, limiting assessment of generalization and bias.
Strategic Implication: If integrated into clinical decision support systems, the framework could reduce manual calculation errors for standardized scores, but its utility is confined to predefined calculators and would require extensive adaptation to cover broader clinical workflows.
Executive Summary: The method improves medical calculation accuracy by grounding LLM outputs in executable tools and constrained extraction. Performance gains stem from better rule grounding, though remaining errors are tied to semantic note understanding and output formatting.
Innovation: 7/10 | Applicability: 6/10 | Commercial Viability: 6/10
PRISMATIC : Prescription Risk Inspection System for Multi-Agent Treatment Interaction in Clinical Decision-Making
Brief: PRISMATIC proposes a multi-agent blackboard architecture where specialized agents collaboratively generate and refine drug regimens for multimorbid patients. It employs two adaptive coordination mechanisms—Dynamic Self-Updating Weight Adjustment (DSWA) and Discrepancy Feedback Precise Calibration (DFPC)—to weigh agent influence and resolve conflicts. Evaluated on MIMIC-IV records, the system shows reduced drug conflicts and improved prescription consistency versus LLM baselines.
Methodological Integrity: The study relies on retrospective MIMIC-IV data, raising concerns about label bias and limited generalizability to prospective clinical settings. No external validation or prospective trial is reported, and the blackboard architecture's scalability to real-time, high-volume prescribing environments remains untested.
Strategic Implication: If integrated into electronic health record workflows, PRISMATIC could assist pharmacists and clinicians in reducing adverse drug events, particularly for complex polypharmacy cases. However, its success hinges on seamless UI integration and clinician trust, which are unproven in the current work.
Executive Summary: PRISMATIC is a multi-agent coordination system designed to improve prescription safety for multimorbid patients using adaptive weighting and feedback mechanisms. Experimental results on MIMIC-IV indicate reduced drug conflicts compared to baseline LLM approaches.
Innovation: 6/10 | Applicability: 6/10 | Commercial Viability: 5/10
PubMed Gems
Human-artificial intelligence collaboration for comanaging epilepsy in real time.
Brief: The platform couples scalp or intracranial EEG signals to a large language model via a smartphone app, enabling bidirectional communication that informs patients of seizure risk, medication adherence, and cognitive state while using patient feedback to retrain the seizure detection algorithm. A pseudoprospective study with 13 inpatient video-EEG participants showed improved detector precision from patient input alone, without expert oversight.
Methodological Integrity: The validation relied on a small, heterogeneous sample (n=13) confined to an inpatient monitoring setting, limiting generalizability to outpatient or real-world use. The pseudoprospective design and lack of a control group raise concerns about bias and the robustness of the reported performance gains.
Strategic Implication: If scaled, the system could reduce false alarms and improve patient engagement in epilepsy management, but its dependence on EEG hardware and active smartphone interaction may hinder adoption compared to ambient, passive solutions. Regulatory clearance and reimbursement pathways for AI‑driven, patient‑facing neurotechnology remain uncertain.
Executive Summary: A prototype EEG‑linked LLM system demonstrated improved seizure detection precision through patient feedback in a limited inpatient cohort. Further real‑world testing and regulatory validation are required to assess broader clinical and commercial viability.
Innovation: 7/10 | Applicability: 6/10 | Commercial Viability: 6/10
AI Clinical Trials (ClinicalTrials.gov)
Venipuncture Via Nurse-AI Collaboration
Brief: The study develops a multimodal deep learning system that fuses visible-light, near-infrared arm images, BMI, and physiological data to reconstruct 3D venous anatomy and suggest optimal puncture sites for contrast-enhanced imaging. Nurses retain final decision authority, accepting, modifying, or rejecting AI recommendations within a standardized human-machine workflow. A multicenter prospective RCT across three hospitals will compare AI-assisted versus routine nursing groups on success rates, procedure time, and safety.
Methodological Integrity: The trial design mitigates bias through randomization and standardized protocols, but potential limitations include modest sample size from only three sites, possible operator variability in image acquisition, and lack of blinding for outcome assessors. Data leakage risk is low as the AI is trained separately, yet external validity may be constrained by the specific imaging equipment and patient populations studied.
Strategic Implication: If proven effective, the tool could reduce venous access failures and procedure delays in radiology departments, offering a modest efficiency gain. However, its impact is limited to a discrete procedural step rather than broader care pathways, limiting its strategic value as a platform technology.
Executive Summary: A multicenter RCT evaluates an AI decision-support tool for peripheral venous access using multimodal imaging and deep learning. The study aims to assess clinical efficacy and human-AI interaction in real-world settings.
Innovation: 6/10 | Applicability: 7/10 | Commercial Viability: 6/10