📡 Frontiers

npj Digital Medicine (published 31 October 2026): Christopher Sauer, Gernot Pucher and colleagues developed and externally validated machine learning models predicting persistent fever 48–72 hours after intravenous antibiotics in hospitalised adults with cancer, using electronic health records from two academic hospitals in Germany and the United States. Across 8,396 hospital stays, the best performing model reached an AUROC of about 0.82. The stated motivation is stewardship: persistent fever commonly triggers extra imaging and broader antimicrobials, so identifying early the patients unlikely to stay febrile could avoid unnecessary escalation.

npj Digital Medicine (30 October 2026): Wayne Han Lee, Lawrence Teck Meng Lam, Sky Wei Chee Koh and colleagues at National University Polyclinics and the National University of Singapore reported MR-AI, a system that scans roughly 25,000 consultation notes a week across seven primary care clinics for referrals a doctor intended but never formally ordered. In a quasi-experimental before-and-after comparison it produced an estimated sevenfold rise in missed referrals detected per week, with 99.9% specificity but 48.8% sensitivity — it missed about half. In a blinded one-week comparison against seven clinicians, 61.8% of its flags were confirmed versus 51.8% for human reviewers.

JAMIA (Volume 33, Issue 11): Cheng Peng, Mengxian Lyu, Ziyi Chen and Yonghui Wu at the University of Florida systematically decomposed clinical data modalities in large language and multimodal models, to quantify how much each modality actually contributes to inpatient risk prediction and decision support. Their conclusion is that the benefit of adding modalities is task-dependent rather than uniformly positive — more input is not automatically better. The University of Florida clinical research institute publicised the work on 26 October 2026, noting that accuracy on the diagnosis and medication tasks remained at or below 0.6.

Annals of Internal Medicine (October 20, 2026): Ben-Gurion University of the Negev and Clalit Health Services reported a benchmark evaluation of six commercial large language models from Google, Anthropic and OpenAI. Tested against Choosing Wisely recommendations, the models generally aligned with evidence-based guidance and rarely recommended unnecessary tests or treatments — but the researchers found they stumbled in clinical situations where human doctors disagree.

The Lancet Digital Health (October 19, 2026): A Mass General Brigham team reported FAHR-Face, a foundation model trained on more than 40 million facial images and fine-tuned into two tools — FAHR-FaceAge, which estimates biological age, and FAHR-FaceSurvival, which stratifies mortality risk in cancer patients from a single photograph. The two rely on different facial regions, and the authors frame the output as decision support alongside a clinician.

Great Wall Cardiology Congress (October 17, 2026): Huawei presented wearable research at the 37th Great Wall Congress of Cardiology in Beijing, including work with Peking University Third Hospital on estimating VO2 max from wearables, and a blood-pressure follow-up analysis drawing on data from over 200,000 wearable users. Speakers positioned continuous monitoring as a move from single measurements toward early warning.

Google (October 12, 2026): Google launched Fitbit Edge, a fitness band priced at $179.99 and positioned between Fitbit Air and Pixel Watch 5. The tracker carries a display 46% larger than Charge 6 in an 11.5 mm aluminium body, and packs GPS plus altimeter, ECG, SpO2, skin temperature and heart-rate sensing — rated IP68 and 5ATM, seven days of battery with the always-on display disabled, and roughly a day of use from a five-minute charge. Orders opened the same day, with shipping from October 27.

Nature Medicine (October 9, 2026): A team led by Shengjing Hospital of China Medical University reported DAMO LiON, a CT-based AI model built to act as a safety net for readers screening tiny liver lesions. In a two-month real-world prospective trial it flagged 15 malignant tumors that had been missed — most of them around one centimetre — allowing earlier surgery or drug treatment, including metastases that are easy to overlook when attention is drawn to a primary tumour.

The Lancet (October 8, 2026): Google and Beth Israel Deaconess Medical Center published the first Lancet main-journal study of AMIE, the conversational diagnostic agent developed at Google. The single-centre prospective feasibility study ran about eight months; 98 adults interacted with AMIE over a secure text channel before an emergency visit, with the system taking a history, discussing likely causes, and generating a visit summary and differential diagnosis for the reviewing clinician.

JMIR Publications (October 1, 2026): Researchers released an educational framework to help psychiatrists assess and discuss patients' use of AI chatbots in mental-health contexts — covering how to screen for reliance on AI for crisis support, how to document AI-related disclosures, and where clinician oversight should begin.

MedTech Spectrum (September 29, 2026): Marani Health received FDA 510(k) clearance for its M•care System (MW-20), which pairs the M•wrap wearable with remote maternal-fetal monitoring. The clearance expands clinician-supervised pregnancy monitoring beyond the clinic, a space where FDA has been scrutinizing AI-assisted remote monitoring claims.

FDA (September 25, 2026): The agency published its first dedicated draft guidance on premarket submissions for robotically-assisted surgical devices (RASD) — Docket No. FDA-2026-N-9505, 91 FR 60967. It covers non-clinical testing, clinical data expectations, and labeling for robots with autonomous or remote-teleoperation capabilities; comments are open until November 24, 2026.

FDA (September 17, 2026): The FDA issued its final decision rejecting a petition to exempt CAD and CADt (AI imaging triage) software from 510(k) premarket notification — the answer is "no." Companies behind AI lesion-detection and triage tools will still need case-by-case 510(k) submissions with performance evidence, keeping the evidentiary bar for new AI imaging devices unchanged.

GI-AI4H (September 16–18, 2026): The Global Initiative on AI for Health — led by WHO and ITU — held its 2026 annual meeting in Hangzhou, gathering health officials from 60+ countries. The agenda centered on AI-enabled multi-disease screening and on moving validated AI health tools from pilots into national primary-care pipelines.

MassDevice (September 16, 2026): Neptune Medical received FDA 510(k) clearance for Triton 1, a robotic flexible endoscopy system designed to give physicians stable control during colonoscopy — reaching anywhere in the colon for both diagnostic and therapeutic procedures, addressing the scope-stability problem that complicates polyp detection and removal.

NIH (September 14, 2026): An NIH-funded UCSF team showed that a single brain-computer interface can decode both speech and upper-body gestures in people with vocal-tract and bodily paralysis — translating brain activity into simultaneous language and gestures to drive a full-body virtual avatar. A concrete step toward more natural communication for people with severe paralysis.

Apple (September 9, 2026): Apple Watch Series 12 debuts the new Health Sensing System — heart rate measured every 5 seconds throughout the day, HRV sampled up to 24× more often, day/night Vitals views, a new 0–10 Readiness score, and a redesigned Health app with Insights, Longevity and a Health Age readout. watchOS 27 ships September 14 with Audio Intelligence (Sound Recognition, Live Rewind, Siri Recap). Apple does not announce any blood-pressure sensing feature; on sale September 18 starting at $399.

Google DeepMind (September 8, 2026): Released the AlphaGenome Atlas — a 1-petabyte, freely searchable catalogue of AI-predicted molecular effects for all 9 billion possible single-letter changes in the human genome, plus more than 100 million short insertions and deletions. A new AVI (AlphaGenome Variant Impact) score separates pathogenic from harmless variants in clinical genome databases; a Broad Institute team used it to flag a non-coding variant as the likely cause of severe epilepsy in one patient. Free for non-commercial research — not cleared for clinical use.

Nature Biotechnology (September 7, 2026): In a 42-patient phase 2a analysis of rentosertib — an AI-discovered TNIK inhibitor — six independently built proteomic aging clocks all showed reduced predicted biological age in treated IPF patients: about 3–4 years at the 30 mg twice-daily dose, up to roughly 6 years on one clock, alongside a dose-dependent improvement in lung function (FVC). The sample is small and the analysis exploratory; the authors acknowledge it cannot yet separate slowed aging from treated lung disease.

Nature (September 2, 2026): A late-life intervention trial — semaglutide started in 20-month-old female mice — extended median lifespan by roughly 12% and outperformed caloric restriction on most aging-clock measures, while improving memory, muscle function and blood-sugar control. NIH-funded; authors note translation to humans remains uncertain, and the benefit pattern resembles caloric restriction rather than something unique to GLP-1 agonism.

FDA (September 2, 2026): Heartvue.ai cleared HeartVueProton (510(k) K260811) — an AI platform that automatically measures ventricular volumes, ejection fraction, blood flow and 2D dimensions from cardiac MRI. The clearance includes a PCCP, letting the algorithm improve within pre-specified bounds without a new 510(k) for each update.

ESC Congress 2026 / British Heart Foundation (September 2026): An Imperial College London model trained on 10.6 million ECGs reads a routine ECG in under two seconds, flagging reduced heart-pumping function (81% accurate in a 61,520-ECG validation set) and aortic stenosis (90%) so patients can be prioritized for echocardiography. The NHS is now testing it on 590 patients across London and Bristol.

OpenAI (September 1, 2026): ChatGPT for Healthcare can now connect to Epic EHR environments — read-only, importing authorized patient information for visit preparation and cross-record summarization, with role-based access, SSO and audit logs under a BAA. A new Healthcare Public Data plugin links 9 official sources including ClinicalTrials.gov, PubMed, DailyMed and CMS Coverage. OpenAI states the tools are not for diagnosis or treatment; EHR integration is unavailable on personal accounts.

JACC (August 2026): The EQUAL randomized trial provides the first randomized controlled evidence on consumer smartwatch AF screening — smartwatch-based screening detected atrial fibrillation at roughly four times the rate of usual care in at-risk adults.

FDA (August 25, 2026): Abbott received De Novo authorization for Libre Duo 10 Day — the first continuous monitor to measure both glucose and ketones in people aged 2+ with diabetes, after Breakthrough Device designation. Continuous ketone tracking is designed to flag diabetic ketoacidosis risk earlier than symptoms alone.

International Journal of Behavioral Nutrition and Physical Activity (August 2026): A systematic review and meta-analysis of randomized controlled trials finds CGM-based glucose feedback improves dietary and physical-activity behavior in people both with and without diabetes — real-time physiological feedback changes behavior.

JACC Advances (August 2026): A systematic review of smartwatch atrial-fibrillation detection finds accuracy varies substantially across devices, with wide differences in false-positive rates — accuracy claims are device- and algorithm-specific, not generalizable.

PLOS Digital Health (August 2026): A systematic analysis of all 1,357 FDA-cleared or approved AI/ML-enabled medical devices finds that only 3 (0.2%) underwent evaluation for patient-centered outcomes, and only 34 had any published clinical evidence — a stark reminder that regulatory clearance is not the same as proven benefit.

Google (August 12, 2026): The new Pixel Watch 5 and Fitbit Health Guardian add blood-pressure pattern tracking and insulin-resistance risk estimation using existing sensor data — moving consumer wearables further toward clinically meaningful cardiovascular-metabolic screening.

Nature Reviews Drug Discovery (August 7, 2026): A comprehensive review of AI in drug discovery — roughly $8.9 billion raised in 2024, yet no AI-discovered or AI-designed drug has received full FDA approval; the authors outline a translational roadmap to close the gap.

FDA clears first Software as a Medical Device with patient-facing LLM: UpDoc Inc received 510(k) clearance (June 25, 2026) for the first AI-powered virtual care platform that generates conversational medical advice directly for patients — a regulatory watershed for LLMs at the patient interface.

Olympus CADDIE EAGLE Trial: AI-assisted colonoscopy achieved a 7.3% absolute increase in adenoma detection rate (ADR) vs. standard colonoscopy — with the strongest gains on high-risk and hard-to-detect colorectal lesions, per multi-center RCT data published February 2026.

European Heart Journal (March 2026): A landmark review maps how wearable devices — smartwatches, rings, CGMs — are capturing longitudinal, real-world physiologic data that now informs clinical decisions across the cardiovascular disease continuum, from early detection to post-discharge monitoring.

JAMA Dermatology (March 2026): A systematic review and meta-analysis finds AI skin-cancer detection systems perform at levels comparable to dermatologists — but real-world generalization across skin tones and devices remains the open challenge.

The Lancet (January 2026): The MASAI randomized controlled trial (>105,000 women) shows AI-supported mammography improves cancer detection and reduces the rate of aggressive diagnoses in the years that follow.

World Health Organization (February 2026): Updated global guidelines on physical activity and sedentary behavior emphasize breaking up sitting time and the value of short, frequent movement bouts.

Nature Digital Medicine (June 2026): AI-powered CBT chatbots demonstrate a measurable reduction in anxiety symptoms (GAD-7, d ≈ 0.52) in a 540-participant randomized controlled trial — not replacing human therapists but serving as an effective bridge to care.

Journal of Medical Internet Research (May 2026): AI productivity tools increase focus time by 23% in controlled settings, primarily by offloading low-level cognitive tasks that would otherwise fragment deep-work sessions.

Harvard Health Publishing (April 2026): An 8-week HRV biofeedback training trial shows significant reductions in cortisol and self-reported stress markers — slow-paced breathing at 5–6 breaths/min emerges as the most accessible evidence-based intervention.

Daraxonrasib (RMC-6236), a new pan-RAS inhibitor, roughly doubles survival in late-stage pancreatic cancer — a disease with a 13% five-year survival rate. Phase 3 trial results published June 2026 mark the first major advance against KRAS-driven pancreatic tumors in decades.

FDA released updated guidance (Jan 2026) for AI-enabled wellness devices and wearables, clarifying the line between "general wellness" products and regulated medical devices. Blood pressure and glucose-monitoring wearables may now enter the market under revised wellness pathways.

Harvard's David Sinclair team received FDA approval for the first human trial of a gene therapy designed to reverse cellular age — using Yamanaka factors to partially reprogram cells. Enrollment began Q2 2026, marking a historic shift from slowing aging to clinically attempting reversal.

Google DeepMind and Isomorphic Labs' AlphaFold3 now predicts structures for nearly all biological molecules, transforming protein design pipelines across pharma. The era of AI-native drug discovery is accelerating.

Patient-specific computational models — "digital twins" — are moving from research labs into hospital pilot programs, with early applications in oncology treatment planning and cardiovascular surgery simulation.

Testing a new continuous glucose monitor + AI dietary recommendation system: real-time blood sugar response to different foods, visualized.

Completed a hands-on review of 5 AI mental health tools: chatbot-based CBT apps, mood trackers, and AI therapy companions. The gap between marketing and clinical rigor remains wide.

Finished reading The Digital Doctor, Chapter 6 — when treatment advice comes from algorithms, how is doctor-patient trust rebuilt? No standard answer, but generative medical narratives are reshaping human connection.

Tried a cuffless blood pressure monitoring prototype: photoplethysmography + AI calibration, consumer devices approaching clinical grade. The biggest challenge is helping users distinguish "physiological fluctuation" from "abnormal values."

Launched the #DigitalSabbath minimalist experiment: 24 hours disconnected from smart devices, tracking time perception and emotional biomarkers with traditional notes. Initial thought — digital withdrawal reactions are more "biological" than expected.

New open-source tool recommendation: Personal Knowledge Graph + biometric correlation analysis — what emerges when your note system can link heart rate fluctuations to creative productivity?