Mount Sinai tested a voice-based LLM assistant across 1,606 pre-procedure calls for cardiac catheterisation, with nurse review and escalation built into the workflow.

Vitestro’s Aletta creates a US route for supervised robotic blood draws: the FDA authorised Aletta through the De Novo pathway, establishing a new device category for autonomous robotic phlebotomy.

GatorOnco is an agentic LLM system for colorectal cancer treatment planning that combines domain adaptation, reinforcement learning and guideline retrieval.

Network Bio launches with $50M and a $30M+ AI collaboration: Network Bio builds disease-specific AI models from human tissue, blood, molecular and longitudinal clinical data; an undisclosed Fortune 100 healthcare company will use the platform to develop next-generation disease-signature models.

TQA and PatientGenie bring patient outreach agents into health system workflows: TQA builds agentic AI integration systems; PatientGenie builds voice and text AI agents for outreach, scheduling, care gap closure and post-discharge follow-up. Deal value not disclosed.

Queen Victoria Hospital introduces Harrison.ai CXR for chest X-ray triage: the NHS trust is using decision-support AI to flag up to 124 chest X-ray findings and prioritise exams for clinician review.

NICE Listens puts public values into the AI governance conversation: NICE’s public dialogue programme is being used to understand how people think about trust, fairness, transparency and human oversight in health and care AI.

EU AI Act delay gives medtech companies more time, not a free pass: high-risk AI Act obligations for AI-enabled medical devices move to 2 August 2028, but transparency, AI literacy and audit preparation still matter now.

Can a voice-based conversational AI assistant safely and reliably support pre-procedure patient preparation for cardiac catheterisation?

Prospective, two-phase implementation at Mount Sinai’s cardiac catheterisation laboratory.

Phase I: 90-day stabilisation period with clinicians and developers improving the workflow.

Phase II: 90-day real-world routine-use period led by nurses.

Sofiya called patients to provide instructions, collect allergy, medication and clinical-status information, answer questions and escalate when needed.

Every AI-generated report was reviewed by a registered nurse before it entered the electronic health record (EHR).

The main outcome was the share of calls that reached the end of the script, with all clinical questions answered.

Phase I included 806 calls. 696 were completed, giving an 86.4% completion rate.

Phase II included 800 calls. 703 were completed, giving an 87.9% completion rate.

Fully automated calls increased from 295 in Phase I to 341 in Phase II.

Protocol-driven nurse callbacks fell from 303 in Phase I to 268 in Phase II.

AI system errors fell from 48 calls in Phase I to 24 calls in Phase II.

Hallucinations that caused incomplete calls occurred in 6 Phase I calls and 0 Phase II calls.

Sofiya-assisted preparation averaged 8.9 minutes, compared with a 20-minute manual baseline.

The authors estimated annual savings equivalent to 37.3 twelve-hour nursing shifts.

Patient satisfaction among surveyed patients was 94.7% in Phase I and 98.1% in Phase II.

This was not autonomous clinical handoff. Nurses still reviewed calls and verified AI-generated reports.

Speech recognition remained vulnerable to background noise, accents, speaking style and vocabulary.

This is the first FDA-authorised standalone robotic blood-draw device; one trained phlebotomist can oversee up to three Aletta devices at the same time.

The Tech: The setup uses near-infrared light, Doppler ultrasound, robotics and AI used to identify suitable veins and guide the draw.

It was developed using 282 billion biomedical tokens, including 166 billion tokens of UF Health clinical text.

It combines domain adaptation, model merging, post-training and agent-based reinforcement learning.

It uses retrieval-augmented generation (RAG) to pull in time-sensitive clinical guideline material during inference.

The linked repository includes code for preprocessing notes, building datasets, constructing guideline-derived vector databases, running inference and evaluating outputs.

The repository reports a blinded evaluation on 79 real-world colorectal cancer cases reviewed by five board-certified oncologists.

The arXiv abstract reports expert-level parity on correctness, currency and safety, with higher scores for readability and completeness versus expert oncologists.

Network Bio launches with $50M to build disease-specific AI models from human tissue data: Network Bio builds AI models trained on patient tissue, paired blood, molecular data and longitudinal clinical outcomes; its $50M financing supports AI-ready biobank datasets, and a $30M+ co-development and licensing deal with an undisclosed Fortune 100 healthcare company will target next-generation disease-signature models.

TQA and PatientGenie integrate AI outreach agents into health system operations: TQA is an agentic AI consultancy that connects AI agents to enterprise systems; PatientGenie is a healthcare AI company whose voice and text agents automate outreach, scheduling, care gap closure and post-discharge follow-up. The partnership targets health plans and health systems. Deal value not disclosed.