Applying the insulin self-management paradigm to heart failure, the SURPASS-HF trial enrolled 21 adults (mean age 69±11 years, 52% women) with implanted pulmonary artery pressure sensors across 90 days. Patients self-executed loop-diuretic dose adjustments guided by daily PA diastolic readings, with every change confirmed by a clinician and supported by ARTHUR, a domain-trained large language model. Time-in-optimal-PA-pressure-range reached 88.4% (intention-to-treat), reading adherence was 92.1%, and no pre-specified safety events — including AKI, hyperkalemia, or symptomatic hypotension — were detected, though the authors candidly acknowledge that sparse lab draws preclude ruling out meaningful harm rates.
The conceptual leap here is significant: pulmonary artery sensors like CardioMEMS have demonstrated mortality and hospitalization benefits in trials like GUIDE-HF, yet the data has remained locked in clinic workflows, limiting scalability across large panels. SURPASS-HF asks whether patients can safely close that loop themselves — a question with real operational urgency given provider shortages in heart failure care.
Critical caveats are substantial: this is a single-arm, 21-patient, single-center feasibility study with no comparator and insufficient statistical power. The cohort was already near target at baseline (PAD 14.8 mmHg), limiting the ability to assess real-world diuretic titration stress. LLM reliability and liability frameworks remain unresolved. As a preprint posted on medRxiv and not yet peer-reviewed, these results — while promising — require validation in randomized, adequately powered trials before any clinical translation.
