OE AI tool outperforms general-purpose models in clinical accuracy and utility

Been critical on OE here and there, but this is the eval data sharing done right. Open-sourced questions/prompts and the actual output response: All available on HuggingFace, open for all others to try and do it themselves! HuggingFace link: https://lnkd.in/gc7kzSPt

Early read of an excellent paper from Jean Feng (UCSF), Anupam Jena, Patrick Heagerty and colleagues. A blinded, specialty-matched evaluation on 620 real point-of-care queries, graded by 149 physicians across 36 states. OpenEvidence, a specialized CDS harness, beats GPT-5.5, Gemini 3.1 Pro & Claude Opus 4.8 on all five axes. Gemini leads among the foundation models, ties Claude on accuracy, and GPT-5.5 places last throughout. A strong reply to the NYU Langone paper (https://lnkd.in/gXsqfuCk), which found the reverse using rubric scoring by 12 raters at one institution. The post below acknowledges this implicitly by saying “for the past two weeks…”. Different method, oposite result. The flip raises methodology questions I am not qualified to decide. I am sure there will be a lot of foundation model vs harness discussion to come: Some EGs 1. Deployment realism. The authors queried the foundation models at temperature 0. That is a lab setting, not what a clinician meets. Frankly it is what deterministic ML scientists do to make GenAI more predictable. We know from recent work (Stanford’s CARE on Arxiv EG) that there are trade offs. Picking one end may well be the wrong choice even if it is the “safe” one. OE ran as a product, settings undisclosed. A second comparison at a realistic temperature would help on both sides. Maybe even three temperatures. The notion that this is an efficient frontier among several variables is already implicit in the paper. 2. One harness, home queries. The questions come from OE’s own platform. The limitations (p 8) and disclosure. (P 13) states OE co-designed the data collection, ran the survey, and paid responders. The authors stayed independent and analysed blind. Fair enough. Why only one harness? A second, say Doximity or UpToDate Expert AI, would separate “OE wins” from “harnesses win.” 3. Reproducibility. The queries are now public on Hugging Face. Generous. But the ratings freeze to specific answers from specific model versions, and OE carries no version number. A snapshot, not a live benchmark. Or rather, if the foundation models improve on future runs it will be criticized as having cheated on the exam 4. Model or harness. The models ran bare, with minimal prompting. OE is a full retrieval and citation stack. Is the finding about the models or the scaffolding around them? If the scaffolding does the work, both papers may be saying the same thing. This harness point is likely to get be important at ex-US installs. Harnesses are based on a limited corpus that is reflective of US practice patterns(NEJM/JAMA EG). Pretty excellent but at a specialty level there may be Improtant differences on how “the literature “ handles different cases and reflect societal past practice (aka biases). By definition, harnessing is “pre-judging “ I think. Grateful to the authors for releasing the data. And for balancing the record. This Foundation vs Harness debate is Important

Clinical AI Smart Fight. Models Lose. Context Wins. Jim Gloats ... Again. I can't judge a smart fight between physicians and the biostatisticians at UCSF, Harvard, and NYU - they apparently prefer confidence intervals to octagons. But here's the story. Last month, authors in Nature Medicine put the leading frontier models (GPT, Claude, Gemini) against specialized clinical tools and they said that the frontier models won. This week, a blinded study sponsored by OpenEvidence flipped it. They said they beat the frontier models across the board. Can they both be right ... or wrong? You don't have to be a genius to spot the funny part: even in the OpenEvidence study, the authors don't credit the win to a better model. They highlight "targeted engineering and customization." They also acknowledge that the frontier models were run through a plain chat box - no pipeline or context. Uh, ok, not exactly a gentlemenly fight. Anyway. Read the two papers side by side (or have your favorite AI do it) and you’ll recognize the question was never "custom vs frontier." If you want great results, you have to own the context. Sound familiar? Like I say - the model is never the moat. Happy 4th! (And apologies in advance to my friends in Britain for this gross mischaracterization at their expense. Harry Kane did it today and made my entire family very happy).

Preference is not correctness. Pairwise evaluation has become a popular way to assess and optimise AI systems for understandable reasons. It is simple, scalable, and people often find it easier to choose between two answers than to assign an absolute score. But in clinical AI, “Which answer do you prefer?” is not always the right question. Both answers may be wrong. The preferred answer may simply be more fluent, confident or comprehensive in appearance. A model can win most comparisons without being calibrated, medically accurate or safe enough for clinical use. This distinction becomes especially important when evaluating objective clinical ground truths such as drug doses, contraindications, physiological variables, epidemiological estimates and predicted risks. These require metrics with interpretable units, reference standards and clinically meaningful error thresholds. Pairwise preference still has an important role but primarily where the outcome is genuinely preference-dependent: patient communication, clinician usability, presentation, tone and workflow integration. My latest article explores why clinical AI evaluation should follow a clearer hierarchy: Correctness and safety first. Preference afterwards. I would be interested to hear how others are separating preference, factuality and clinical safety in medical AI evaluation. https://lnkd.in/exfpwq4S #ClinicalAI #MedicalAI #ArtificialIntelligence #MachineLearning #AIAlignment #RLHF #HealthcareInnovation #PatientSafety #AIEvaluation #PairwisePreference

Still considering deploying an LLM-based clinical AI solution? “Although many models achieve high scores on established benchmarks, our adversarial stress tests demonstrate that this knowledge is often brittle and unsafe in practice. This gap manifests across all four safety axes we investigated. In our robustness audit, a median of 94% of initially correct answers in MedQA failed under dynamic perturbations” A prime proof of textual retrieval in place of understanding.

A new peer-reviewed study in Nature Portfolio Medicine tested two of the most widely deployed specialized clinical AI tools — OpenEvidence and UpToDate Expert AI — against general-purpose frontier LLMs (GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6). The frontier models won. Across every benchmark — medical knowledge, clinician alignment, and real-world clinical queries — general-purpose LLMs outperformed the domain-specific tools. The specialized clinical AI tools performed no better than Google Search AI Overview. The study's conclusion is pointed: "These findings highlight the need for independent, real-world evaluation of AI tools before they enter clinical settings." What this study doesn't address, and what we think matters most, is what happens after the query is answered. Getting the right response is only half the equation. The AI still needs clean, clinician-verified, structured data to work with. A smarter model pointed at fragmented records doesn't fix the underlying problem. That's what Medicomp is built to solve. https://lnkd.in/eXJ5asyv 

One theme continues to emerge across healthcare organizations implementing AI: The most successful initiatives rarely begin with AI. They begin with an operational problem. • Reduce no-shows • Improve access • Decrease administrative burden • Shorten documentation time • Improve workforce efficiency Then AI becomes an accelerator—not the strategy itself. Too many organizations still ask: "Where can we use AI?" High-performing organizations ask: "What business outcome are we trying to improve?" Strategy first. Workflow second. Technology third. That's how transformation becomes sustainable. #HealthcareOperations #AI #DigitalTransformation #ExecutiveLeadership #StrategyExecution https://lnkd.in/geUVH65Y

1. Refusals weren't discarded as noise, they were a reported endpoint. UpToDate's 19% refusal rate (vs. 1–6% for others) is in the paper as its own figure, statistically tested. It is excluded from the headline comparison score, so that critique holds for the number going viral. But the data cut against "refusal = safer": harmful-content and hallucination rates didn't differ significantly across any of the six systems (Cochran's Q, P=0.55 and 0.42). 2. I certainly agree with you that the query-source bias is your strongest point, and it's real. The 100 RCQ queries came from a log of physicians already using a general-purpose chatbot. That's a selection effect baked into the design: anyone who wanted a fast guideline citation may have gone straight to OpenEvidence/UpToDate and never generated a query in this sample. The authors (unfortunately) don't address this (for some reason(s)). 3. Circularity, already conceded by the authors, more strongly than implied. The HealthBench judge panel was GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6, the exact three systems being scored. The authors acknowledge this and, because of it, explicitly demote HealthBench to "supplementary," naming the blinded 12-clinician RCQ review as primary evidence instead. There's a second circularity layer too: HealthBench is OpenAI's own benchmark, and GPT-5.2 scored highest on it and also flagged by the authors as a possible benchmark-developer overlap. One nuance missing from this post: the gap is concentrated in completeness/clarity, not core knowledge. Across dimensions, models differed least on clinical correctness (Kendall's W=0.14) and most on clarity (W=0.29), in that sense, OpenEvidence's main weakness was communication, not accuracy. "Specialized tools know less medicine" overstates what the data show; "specialized tools organize and complete answers less well" is the more defensible read. That's why the authors themselves frame this as a snapshot, not a verdict and that framing matters beyond the methods debate. Headlines compressed to "general AI beats specialized clinical tools" can travel further than the caveats attached to them. Lay readers, patients, or even health-system buyers skimming the summary may walk away thinking the matter is settled, when the underlying data show a narrower and more conditional finding: a gap concentrated in completeness and clarity, not core medical correctness, measured on one slice of real-world queries with real limitations. Treating a single brief communication as definitive risks two kinds of harm, individuals leaning on general-purpose chatbots for decisions that warrant clinician-vetted, guideline-anchored tools; and health systems making procurement or investment calls that sideline retrieval-based safeguards before the evidence is mature enough to justify it. Either way, the people most exposed to that overconfidence are the patients whose care depends on getting it right.

Clinical AI does not operate without bias. Every model reflects assumptions about the data it uses, the outcomes it prioritizes and the patients it is designed to serve. Writing in MedCity News, Shakira J. Grant MD, MSCR, a member of the standards committee that helped develop URAC’s Health Care AI Accreditation, explains why the real question is not whether bias exists, but whether health care organizations understand, document, test and govern the assumptions behind their AI tools. Read more: https://hubs.la/Q04n3s1P0 #AI #HeatlhAI #HealthcareAI

People often think the output of an #AI system is a prediction. In #healthcare, the real output is trust. Because a decision accompanied by traceable evidence, reproducible logic, and transparent reasoning becomes something clinicians can evaluate with confidence. This week at #UMMS, I am presenting my work on a particular #clinical decision problem and how I went from ambiguity to a traceable decision pipeline.