Most AI alerts in a heart-failure trial prompted a plan to check again

In LINK-HF2, 71 of 108 AI notifications communicated to clinicians led to a plan to reassess the patient’s symptoms three days later. That planned follow-up was the most common response. The trial did not establish a reduction in heart-failure readmissions.[1]

Published in JACC: Heart Failure on 9 October, the study randomised 176 patients at five US Veterans Affairs centres and analysed 171. Both groups wore sensors for up to 90 days after hospital discharge; clinicians received AI notifications only for the intervention group.[1]

What the warning asked a clinician to do

The registered protocol describes a chest patch linked to a smartphone, with measurements uploaded automatically for analysis. A heart-failure nurse would review monitoring information alongside the patient’s symptoms and blood pressure, then assess a response with a physician. The available routes included adjusting medication, arranging a clinic visit or referring the patient to an emergency department.[2]

The underlying approach is described in the separately published 2024 pilot study: machine learning builds a personal model from early measurements, then looks for changes in heart rate, breathing and activity. The alert brings a change to the team’s attention; establishing what caused it and what would help remains a clinical task.[3]

A departure from someone’s usual physiology contains less information than a treatment decision requires. A clinician has to connect the signal to that person’s present condition, medication and recent history. Faster recognition makes this decision possible earlier; it does not supply all the missing clinical information.

The hospital outcome

Clinical outcomes were exploratory, and the wide confidence interval leaves a substantial range of possible effects.[1]

Why another check was an intentional option

The earlier pilot had already exposed the work behind an alert. In five observed encounters, clinicians reviewed records for three to 15 minutes and spent 12 to 22 minutes on patient calls. They examined symptoms, medications and other information before choosing a response.[3]

That early experience changed the trial: researchers reduced repeated notifications, expanded the information available to clinicians and added an increased-monitoring option. Reassessment was a deliberate response to the pilot, not simply an absence of action.[3]

LINK-HF2 therefore points to a demanding design problem for medical AI: make a warning useful at the moment a clinician must choose what happens next. The prospect is earlier care, but achieving it requires a clear connection between a detectable change and an intervention likely to help that particular patient.

Report an error or suggest a correction