LLMs outperform clinical AI in landmark study; new data from sleep, oncology, and eye disease
Nature Medicine benchmarks, a JAMA eye-screening trial, ASCO oncology data, and MHRA's new AI drug safety programme.
By Dr. Asher Knippel
Today's roundup covers seven stories from the last 72 hours.
Wednesday, 17 June: General-purpose AI outperforms dedicated clinical tools in Nature Medicine
A study published in Nature Medicine on 12 June 2026 delivered a striking benchmark result: three general-purpose large language models — OpenAI's GPT-5.2, Google's Gemini 3.1 Pro, and Anthropic's Claude Opus 4.6 — outperformed two specialist clinical AI products, OpenEvidence and Wolters Kluwer's UpToDate Expert AI, across every benchmark tested. Researchers evaluated the systems on 500 MedQA clinical-knowledge questions, 500 HealthBench items measuring alignment with clinician responses, and 100 real patient queries drawn from NYU Langone Health. Gemini scored 97.4% accuracy on MedQA, against 89.6% for OpenEvidence and 88.4% for UpToDate. The authors note that healthcare-specific branding and curated retrieval do not inherently confer a clinical performance edge over frontier general models. The finding does not resolve the separate questions of regulatory approval, liability, or clinical workflow integration — but it does complicate the assumption that purpose-built clinical products are the only safe default.
Wednesday, 17 June: AI eye-screening tool cuts false-positive diabetic referrals by 45% in JAMA trial
A two-part study published in JAMA on 15 June 2026 shows that an AI-powered optical coherence tomography (AI-OCT) system can dramatically reduce the burden of unnecessary referrals from diabetic macular oedema (DME) screening — without missing genuine disease. In a prospective validation of 603 patients, the system achieved 98.8% sensitivity and 90.7% specificity for detecting DME. A subsequent multicentre, non-inferiority randomised clinical trial in 276 patients then demonstrated a 45% reduction in false-positive referrals compared with standard fundus photography alone, with zero missed cases. Diabetic macular oedema is the leading cause of preventable vision loss in working-age adults with diabetes; streamlining referrals without sacrificing sensitivity is a meaningful clinical gain, particularly in settings where specialist appointments are scarce.
Wednesday, 17 June: SLEEP 2026 — Phase 3 narcolepsy cognition data and new solriamfetol analysis
At the SLEEP 2026 annual meeting in Baltimore (14–17 June), Axsome Therapeutics presented data from two narcolepsy studies. Results from a Phase 3 trial of AXS-12 (reboxetine) examined cognitive and functional outcomes in patients with narcolepsy with cataplexy — a dimension historically underweighted in wakefulness-focused narcolepsy trials. Axsome also presented a post-hoc analysis of solriamfetol (Sunosi) for excessive daytime sleepiness across narcolepsy and obstructive sleep apnoea subgroups. Narcolepsy affects an estimated 3 million people globally and is significantly underdiagnosed; data on cognitive endpoints could strengthen the case for broader prescribing and support updated labelling.
Wednesday, 17 June: UK medicines regulator launches AI sandbox to catch adverse drug reactions earlier
The UK's Medicines and Healthcare products Regulatory Agency (MHRA) announced on 15 June 2026 the creation of a regulatory testing environment — termed a "sandbox" — to evaluate whether AI can flag medicine safety concerns before they cause large-scale harm. The programme will bring together regulators, researchers, and technology developers to test AI-assisted pharmacovigilance tools in a controlled setting. The motivation is concrete: an estimated 250,000 hospital admissions occur annually in the United Kingdom as a direct result of adverse drug reactions that conventional post-marketing surveillance failed to catch in time. Positive results from the sandbox could lead the MHRA to incorporate AI into its standard regulatory toolkit — and could serve as a model for other European medicines agencies.
Wednesday, 17 June: ASCO 2026 — Translation is now precision oncology's hardest problem
The 2026 American Society of Clinical Oncology (ASCO) Annual Meeting, which concluded on 15 June in Chicago, delivered a consistent message: the hard part of precision oncology has shifted. Discovering new biomarkers and targeted agents is no longer the primary bottleneck — bringing them to patients at scale is. Across presentations on targeted therapies, liquid biopsy, molecular diagnostics, and AI-assisted pathology, speakers returned to a shared obstacle: tools that work well in academic centres stall at the point of integration into routine clinical workflows. The meeting's theme, "The Science and Practice of Translation," named this explicitly. For healthcare systems in Cyprus and the eastern Mediterranean, where access to high-cost molecular profiling may be constrained, the drive towards cheaper, deployable alternatives is directly relevant.
Wednesday, 17 June: NIH funds $6 million programme to screen 100 billion drug candidates for Alzheimer's
Indiana University School of Medicine and the IU Luddy School of Informatics have launched a five-year Alzheimer's drug discovery programme, backed by a $6 million grant from the US National Institutes of Health (NIH). The project will use AI and machine learning to scan chemical libraries of up to one hundred billion molecules, ranking candidates most likely to interact with proteins implicated in Alzheimer's disease and related dementias. Conventional computational screening of comparable chemical space takes years; the AI platform compresses this into minutes. The work runs alongside IU's TREAT-AD programme. This is early-stage, preclinical research: the programme is identifying candidates, not yet testing them in human participants. If promising compounds emerge, they would face years of safety and efficacy trials before any clinical use.
Wednesday, 17 June: Novel CNS drug SUVN-I6107 advances from Phase 1 to Phase 2
Suven Life Sciences announced on 16 June 2026 that SUVN-I6107 — a muscarinic M1 positive allosteric modulator (M1-PAM) — has successfully completed its Phase 1 first-in-human study and will progress to Phase 2. The Phase 1 trial confirmed a favourable safety and tolerability profile, with no dose-limiting toxicities observed. Pharmacodynamic measurements verified that the compound engages its target in the central nervous system. Muscarinic M1 receptors play a key role in cortical and hippocampal cognition; M1-PAMs are under active investigation for Alzheimer's-related cognitive decline and, separately, for schizophrenia-linked cognitive deficits. This is Phase 2 entry: efficacy in human patients remains to be established. The Phase 2 trial design has not yet been publicly announced.
This article is for informational and journalistic purposes only and does not constitute medical advice. Readers should consult a qualified clinician before making any change to their treatment, medication, or health management plan.