TOPLINE

In patients recently discharged after a hospital stay for heart failure (HF), AI-guided monitoring with clinician alerts appeared feasible, but it did not reduce rehospitalizations or the composite of rehospitalizations, emergency visits, and death.

METHODOLOGY

• Researchers conducted a prospective, randomized, open-label study (LINK-HF2) at five US Veterans Affairs medical centers to assess whether pairing AI-generated alerts with a structured clinician response plan was feasible for care after a hospital stay for HF.
• They enrolled adults being discharged after hospitalization for HF between January 2022 and May 2024. The final analysis included 171 patients (mean age, 72.6 years; 93% men).
• Patients were randomly assigned to the intervention group (n = 80) or the control group (n = 91). All patients wore a chest sensor patch for up to 90 days, but only clinicians of those in the intervention group received AI alerts.
• The patch tracked each patient's heart rate, movement, and skin temperature, and the AI set a personal baseline within 48 hours of discharge. It raised an alert when readings deviated from baseline values, and clinicians contacted the patient and followed one of four response pathways, ranging from treatment adjustment to emergency department referral.
• Patients were followed up for a median of 90 days. HF rehospitalization was the main exploratory clinical endpoint; researchers also assessed feasibility and implementation metrics and a composite of all-cause unplanned rehospitalizations, emergency department visits, and all-cause mortality.

TAKEAWAY

• Patients wore the sensor for a median of 72.4% of follow-up days, and clinicians responded to alerts within a median of 1 day. Most responses (65.7%) involved a symptom recheck in 3 days, and no device-related serious adverse events occurred.
• The risk for HF rehospitalization did not differ significantly between the intervention and control groups.
• The composite clinical endpoint of all-cause unplanned rehospitalizations, emergency department visits, and all-cause mortality also showed no significant difference between the groups.
• In a post hoc analysis, the intervention group had a lower rate of AI-detected physiologic instability than the control group (hazard ratio, 0.76; P = .025), but the difference was not significant when death was treated as a competing risk.

IN PRACTICE

"The central insight from LINK-HF2 is that in AI-enabled medicine, accurate physiological risk detection alone does not modify disease trajectory unless it is coupled with consistent mechanistically aligned and trusted therapeutic responses," the researchers wrote.

SOURCE

The study was led by Konstantinos Sideris, MD, of George E. Wahlen Department of Veterans Affairs Medical Center in Salt Lake City. It was published online on October 9 in JACC: Heart Failure.

LIMITATIONS

The study had a small sample size and a short follow-up duration. Most patients were older men treated in one health system. Baselines were set within 48 hours of discharge, which may have been inaccurate if patients were not yet stable.

DISCLOSURES

The study received funding from the Veterans Health Administration Health Services Research and Development Service. Several authors reported being employees of Prolaio, the operational partner that provided the sensors, analytics platform, and clinician portal. Two other authors reported receiving consulting fees, and one of them also reported research support, from multiple health care companies and organizations. Detailed disclosures are reported in the original article.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.