Scope: Mar 07 – Mar 14, 2026

Headline Insight: The PCP in Your Pocket

This paper is less a research curiosity and more a milestone marker. Google and Beth Israel have done something the field has been waiting for: taken a conversational diagnostic AI off the simulation track and into a real ambulatory clinic, with real patients, real physicians reviewing the output, and real chart review eight weeks out to validate diagnoses.

The headline numbers hold up under scrutiny. AMIE matched the final diagnosis in 90% of cases within its top-7 differential, achieved 75% top-3 accuracy, and produced management plan quality statistically indistinguishable from PCPs on appropriateness and safety — the two dimensions that matter most for patient harm. PCPs outperformed on practicality and cost-effectiveness, which is expected: a physician who just examined a patient and reviewed their chart will naturally write a tighter, more contextually grounded workup than a chatbot with no EHR access and no physical exam.

The safety result is the most clinically meaningful number in the paper: zero safety stops across 100 real patient interactions, monitored in real time by board-certified internists. No emotional crises, no clinical harm triggers, no patient asking to stop.

What this actually means for the "do we need more PCPs" debate:
The traditional argument — that primary care shortages require training more physicians — assumes that the bottleneck is fundamentally a supply problem requiring a supply-side solution. This paper quietly challenges that frame. AMIE, running on Gemini 2.5 with no EHR access and no physical exam, produced diagnostic and management quality on par with a PCP for an urgent care complaint in 90% of cases. It did so in a text chat, before the appointment even started, and it handed the physician a structured summary — often exceeding the detail of standard intake documentation.

The limitations the authors flag are real and should not be glossed over. This was 100 patients at a single academic center, English-speaking, with laptops, pre-screened to exclude pregnancy and mental health complaints, and monitored by physicians the entire time. The Hawthorne effect almost certainly suppressed adversarial behavior. The AI had no EHR access. The physical exam — which PCPs used to sharpen their management plans — was unavailable to AMIE, which is why its workups skewed broader and less cost-effective.

But the trajectory is clear. EHR integration is an engineering problem, not a scientific one. Multimodal input — voice, video, even limited visual exam — is already in Google's roadmap for this system. Mental health and pregnancy exclusions reflect IRB conservatism for a first-in-human study, not fundamental model limitations.

The argument that we need more PCPs is not wrong. But this paper introduces a credible alternative frame: we may need PCPs to do different things — complex judgment, procedural care, the relationship and trust elements that patients in this study explicitly said they still wanted from a human — while AI handles the high-volume, pattern-recognition-heavy, time-sensitive intake work that currently saturates primary care capacity.

The PCP shortage crisis was always partly a mismatch between where physician cognitive effort is directed and where it adds irreplaceable value. This paper is early evidence that the mismatch is addressable without waiting 10 years for more graduates.

Pre-Print Intelligence (arXiv)

OncoAgent: Guideline-Aware AI Agent for Zero-Shot CTV Auto-Delineation

Source: arXiv:2603.09448v1 (preprint, not peer-reviewed) | Authors: Kim et al., Oncosoft Inc. & Samsung Medical Center | COI: Four authors are Oncosoft Inc. employees; senior author is CEO

Summary OncoAgent is a two-phase LLM-based agent (planning + execution) that converts free-text radiotherapy guidelines into 3D clinical target volume (CTV) contours without task-specific training data. The framework uses GPT-5.2 as its reasoning engine, pre-trained OAR segmentation models, and geometric operations (dilate, subtract, union) to produce CTV/PTV contours. Evaluated on mid-thoracic esophageal cancer, it achieved zero-shot DSC of 0.842 (CTV) and 0.880 (PTV), statistically non-inferior to a supervised nnU-Net (GTV Prior) baseline.

Methodological Integrity The test cohort is critically small (n=8), drawn from a single institution, with the clinical preference survey derived from only two physicians rating those same 8 cases — insufficient statistical power to support the strength of the clinical preference claims made. The supervised baselines trained on 32 cases are possibly underpowered relative to production-grade DL models, potentially flattering the zero-shot comparison.

Strategic Implication The genuine differentiation is guideline adaptability, not raw segmentation accuracy — this directly addresses a documented friction point in clinical AI deployment where model retraining costs erode ROI after guideline updates. Adoption will depend on robustness of the upstream OAR segmentation layer and regulatory tolerance for LLM-generated treatment planning logic, neither of which is resolved here.

Executive Summary OncoAgent demonstrates proof-of-concept that LLM-based reasoning over clinical guidelines can approximate supervised segmentation performance on CTV delineation without annotated training data, with physician-rated superiority on compliance metrics. The work is preliminary — single-institution, eight-patient test set — and requires multi-center prospective validation, but has high potential.

Sentinel: LLM-Based Autonomous Triage Agent for Remote Patient Monitoring

Source:[arXiv:2603.09052] (preprint, not peer-reviewed) | Authors: Kim et al., AnsibleHealth Inc. | COI: Senior author (Po) is founder and CEO of AnsibleHealth; study conducted entirely on AnsibleHealth proprietary data; manuscript drafting AI-assisted

Summary Sentinel is a production-deployed autonomous triage agent using Claude Opus 4.6 via MCP, equipped with 21 structured clinical tools for dynamic EHR context retrieval, that classifies RPM vital sign readings into four severity tiers. Evaluated retrospectively on 500 readings from 340 polychronic patients across 25 U.S. states, it achieved 95.8% emergency sensitivity, a quadratic-weighted kappa of 0.778 against a six-clinician majority-vote reference, and outperformed every individual clinician on emergency sensitivity (97.5% vs. 60.0% aggregate) in LOO analysis, at $0.34 per triage.

Methodological Integrity The study is entirely single-organization, retrospective, and self-funded by the company whose system is being evaluated — a combination that structurally limits independence of the reference standard. The majority-vote reference itself had clinician unanimity on only 42.8% of samples, creating a ceiling problem that simultaneously inflates apparent overtriage and obscures genuine agent error; the LOO analysis partially corrects this but cannot resolve the absence of ground-truth outcome data.

Strategic Implication The $0.34-per-triage cost structure is the most commercially significant finding — at scale it makes 24/7 contextualized triage economically viable where human staffing is not, directly addressing the documented mechanism of failure in Tele-HF, BEAT-HF, and TIM-HF1. Adoption will hinge on prospective outcome data and payer willingness to reimburse AI-generated triage, neither of which is established.

Executive Summary Sentinel demonstrates retrospective technical validity for LLM-based contextual RPM triage at a cost point that is operationally disruptive, but the single-organization design, CEO-founder conflict of interest, and absence of clinical outcome data leave the core efficacy claim — that this architecture replicates TIM-HF2's mortality benefit — entirely unproven.

FetalAgents: Multi-Agent System for Fetal Ultrasound Analysis

Source: arXiv:2603.09733v1 (preprint, not peer-reviewed) | Authors: Hu, Huang, Liu et al., Tsinghua University & West China Second University Hospital | COI: None declared | Article type: Original research — system paper with external validation

Summary FetalAgents is a multi-agent system built on AutoGen that orchestrates ensembles of task-specific vision models (FetalCLIP, nnU-Net, SAMUS, SAM-based variants) via a GPT-5-mini coordinator for comprehensive fetal ultrasound analysis across eight clinical tasks — plane classification, anatomical segmentation, and fetal biometry — with an additional pipeline for automated keyframe extraction and structured clinical report generation from continuous video streams. Evaluated on multiple independent external datasets, it consistently outperforms all standalone baselines across metrics.

Methodological Integrity External validation datasets are small for several tasks (64 cases for AoP, 75 for HC, 233 for standard plane classification), and performance gains over the strongest baselines are incremental in most categories rather than clinically transformative; the video summarization evaluation is qualitative only, with no quantitative comparison against ground truth keyframe selection or report quality metrics.

Strategic Implication The workflow-to-report pipeline directly addresses the documentation burden in obstetric sonography — a known staffing bottleneck globally — but clinical deployment requires prospective validation, regulatory clearance per task, and integration with existing PACS/RIS infrastructure, none of which is addressed.

Executive Summary FetalAgents demonstrates measurable performance gains over state-of-the-art fetal ultrasound models through agentic orchestration and model ensembling across eight tasks on external datasets, with an end-to-end video reporting capability that has no quantitative validation; the system requires prospective clinical trials and regulatory engagement before deployment.

Meissa: Multi-modal Medical Agentic Intelligence

Source: arXiv:2603.09018v1 (preprint, not peer-reviewed) | Authors: Chen, Bai, Pan, Zhou, Yuille — Johns Hopkins University & Cornell University | COI: None declared | Article type: Original research — system paper with benchmark evaluation

Summary Meissa is a 4B-parameter multimodal medical LLM trained via SFT on ~40K trajectories distilled from Gemini-3-flash, designed to replicate frontier agentic behaviors — tool calling, interleaved image reasoning, multi-agent debate, and clinical simulation — in a fully offline deployment. Training employs a three-tier stratified curriculum that routes queries by difficulty and pairs prospective execution traces with retrospective hindsight re-narrations. Evaluated across 13 benchmarks spanning radiology, pathology, and clinical reasoning, it matches or exceeds frontier models in 10 of 16 settings at 25× fewer parameters and ~22× lower latency.

Methodological Integrity The NEJM OOD benchmark contains only 15 test cases, making results there statistically unreliable; trajectory quality auditing relies on Claude Opus as evaluator — a non-independent gate given Claude's presence in the broader benchmark ecosystem. No prospective clinical validation, clinician-in-the-loop assessment, or image-level OOD contamination audit beyond text n-gram decontamination is reported.

Strategic Implication Offline deployment directly addresses the primary structural barrier to clinical AI adoption — patient data sovereignty and on-premise compute requirements — but the absence of calibrated uncertainty estimation, an abstention mechanism, and any regulatory engagement means FDA SaMD clearance remains multiple development cycles away.

Executive Summary Meissa demonstrates that frontier-level medical agentic reasoning is distillable into a compact, privacy-preserving model competitive with GPT-4o across radiology and pathology benchmarks at a fraction of inference cost; clinical deployment readiness requires prospective validation and regulatory development not yet initiated.

PET-F2I: A Comprehensive Benchmark and Parameter-Efficient Fine-Tuning of LLMs for PET/CT Report Impression Generation

Source: arXiv:2603.10560v1 (preprint, not peer-reviewed) | Authors: Liu, Zhang, Zhang et al. — Fudan University, Shanghai Universal Medical Imaging Diagnostic Center, Lanzhou University, SIAIS | COI: None declared | Article type: Original research — benchmark paper with domain adaptation modeling

Summary PET-F2I-41K is a 41,191-report benchmark drawn from a single Chinese institution (2013–2023) for the task of generating diagnostic impressions from PET/CT findings text, accompanied by three novel clinical metrics — Entity Coverage Rate (ECR), Uncovered Entity Rate (UER), and Format Compliance Rate (FCR) — designed to capture omission and hallucination failures invisible to standard NLG metrics. A domain-adapted 7B model (PET-F2I-7B), fine-tuned from Qwen2.5-7B via LoRA on 40,691 training reports, achieves BLEU-4 of 0.708 and ECR of 0.807, substantially outperforming all 27 zero-shot baselines including frontier proprietary models. The benchmark and model together constitute the first standardized evaluation infrastructure for this subspecialty task.

Methodological Integrity The dataset originates from a single Chinese center, making generalizability to Western institutional terminology, report structures, and tracer protocols undemonstrated; 92.1% of reports are FDG-oncology, rendering minority-tracer results statistically thin despite the presented cross-tracer heatmaps. The clinical metrics rely on a greedy dictionary-match NER framework rather than a validated clinical NLP pipeline, and no radiologist preference study or downstream clinical outcome validation is reported — ECR improvement does not establish diagnostic safety.

Strategic Implication The offline-deployable architecture directly addresses PHI compliance barriers that make cloud-based LLM deployment untenable in most health systems, and the benchmark fills a genuine void in PET/CT reporting automation; however, single-center Chinese-language derivation substantially limits near-term adoption in Western markets without multicenter replication and regulatory engagement.

Executive Summary PET-F2I-41K establishes the first large-scale benchmark and clinically grounded evaluation framework for PET/CT impression generation, with a locally deployable 7B model achieving a 3× entity coverage improvement over frontier zero-shot baselines; clinical deployment readiness requires multicenter external validation, radiologist outcome studies, and regulatory submissions not yet initiated.

Peer-Reviewed Breakthroughs

Clinical Environment Simulator (CES): A Framework for Dynamic Clinical LLM Evaluation

Source: Nature Medicine Perspective, peer-reviewed | Authors: Luo, Kim et al., Harvard Medical School | COI: One author is Google Research employee; one is Microsoft Research; one is cofounder of Capacity Health (AI clinical decision support); one is a visiting researcher at Google DeepMind | Article type: Perspective — conceptual framework proposal, not original empirical research

Summary CES is a proposed dual-engine simulation architecture for evaluating clinical LLMs in dynamic hospital environments, comprising a "hospital engine" tracking real-time resource states and a "patient engine" modeling disease progression in response to LLM interventions. Unlike existing static benchmarks (MedMCQA, PubMedQA) or sequential diagnosis frameworks, CES would score LLMs on both clinical outcomes and operational efficiency metrics across temporal, resource-constrained, and adversarially stressed scenarios. No implementation is described and no empirical results are presented.

Methodological Integrity This is a conceptual perspective with no experimental data, prototype implementation, or validation evidence — the framework exists entirely as a proposal. The core technical challenges the authors themselves acknowledge — simulation fidelity, physiological model accuracy, EHR grounding, and error propagation across coordinated engines — remain entirely unaddressed.

Strategic Implication The framework directly targets a documented gap: current benchmarks cannot assess whether LLMs make temporally coherent, resource-aware decisions, which is the exact capability required for deployment in high-acuity clinical environments. Practical value is contingent on construction and validation of the simulator, which is a substantial, multi-year engineering undertaking.

Executive Summary CES articulates a well-reasoned conceptual direction for next-generation clinical LLM evaluation, published in a high-impact venue by a credible author group, but delivers no implementation, data, or empirical validation — its value is entirely prospective.

AI Clinical Trials (ClinicalTrials.gov)

ENABLE-HCM: AI-Guided Echocardiography by Non-Sonographers in HOCM — NCT07155434

Source: ClinicalTrials.gov registration | Sponsor: UltraSight (industry) | Status: TERMINATED — sponsor cited internal review of program priorities | PI: Milind Desai MD, Cleveland Clinic | No results posted

Summary ENABLE-HCM was a prospective observational cohort study (n=36 actual enrollment) evaluating whether non-sonographer healthcare providers using UltraSight's ML-based guidance software could acquire limited transthoracic echocardiography images of diagnostic quality in HOCM patients eligible for mavacamten, with same-day blinded expert cardiologist comparison against trained sonographer acquisitions. Primary endpoint was LVEF agreement via Simpson's Method; secondary endpoints included view-specific image quality on the 5-point ACEP scale and LVOT gradient classification. The study enrolled at a single site (Cleveland Clinic) and ran from July 2025 to February 2026.

Methodological Integrity With 36 actual enrollees at a single tertiary referral center and no results posted, no conclusions on diagnostic equivalence can be drawn; the study is underpowered by design relative to FDA-grade equivalence thresholds, and the highly selected HOCM population at a center of excellence limits generalizability to community settings where the access argument is most relevant. Termination for "internal review of program priorities" — rather than safety or futility — introduces ambiguity about whether enrollment, data quality, or commercial strategy drove the decision.

Strategic Implication UltraSight's core thesis — democratizing point-of-care echo acquisition through AI guidance — remains commercially viable and technically plausible, but this trial's termination without results leaves the HOCM/mavacamten monitoring use case unvalidated, a meaningful gap given that Camzyos REMS requirements create a defined, recurring need for LVEF surveillance that would directly benefit from non-sonographer acquisition capability.

Executive Summary ENABLE-HCM terminated early with 36 patients and no published results, leaving the primary hypothesis unconfirmed; the strategic rationale — reducing REMS-mandated echo burden in mavacamten patients through AI-guided non-expert acquisition — remains commercially relevant but requires a completed, adequately powered trial to support regulatory or payer claims.

Legal & Regulatory Disclaimer

No Medical Advice: The intelligence, research, and analysis published by FC-OE are strictly for B2B informational and educational purposes. Although the founder is a licensed medical professional, no content provided on this platform, in newsletters, or during advisory sessions constitutes medical advice, clinical diagnosis, or treatment recommendations. Engaging with FC-OE does not establish a doctor-patient relationship.

No Investment Advice: FC-OE provides technical and clinical evaluations of medical artificial intelligence, Software as a Medical Device (SaMD), and digital health technologies. This information does not constitute formal financial, investment, legal, or regulatory advice. Venture capital, private equity, and startup investments carry inherent high risks. Subscribers, funds, and clients are solely responsible for conducting their own independent due diligence and consulting with their own financial and legal advisors prior to making any capital allocation or investment decisions.

Institutional Affiliations: The views, analyses, and commercial viability scores expressed by FC-OE are entirely independent. They do not represent, reflect, or imply the official policy, position, or endorsement of any academic, clinical, or corporate institution with which the founder may be affiliated.

Accuracy & Liability: The fields of machine learning, artificial intelligence, and clinical medicine are rapidly evolving. While FC-OE utilizes rigorous methodologies to curate and evaluate data, we make no warranties or representations regarding the absolute accuracy, completeness, or real-time validity of the information provided. FC-OE and its founder assume no liability for any commercial, clinical, or financial actions taken in reliance upon this publication.