Conversational AI for remote monitoring in heart failure: a prospective controlled pilot study

Heart failure (HF) requires scalable strategies to detect decompensation early and reduce hospitalizations. Existing telemonitoring tools are often invasive, complex, or poorly integrated into routine care. To evaluate the feasibility and clinical impact of a conversational artificial intelligence system (CAIS) for automated telephone follow-up in patients with HF.

We conducted a prospective, non-randomized, controlled feasibility study at a tertiary hospital. Eighty-six outpatients received weekly AI-assisted follow-up via natural language processing calls, collecting symptoms and vital signs. Alerts were reviewed by HF nurses, who determined responses per standard practice. Forty patients received usual care. The primary objective was feasibility and acceptability; exploratory endpoints included all-cause death, cardiovascular death, HF hospitalization, and diuretic intensification at 12 months, analysed with Cox and competing-risks regression. Of 4272 scheduled calls, 3919 were completed (91.7%). CAIS generated 1962 alerts, prompting 648 actions-mainly nurse calls (86.6%) and medication changes (7.9%). Nurse workload was 2.4 min/patient/week. At 12 months, the CAIS group improved KCCQ-12 score by +7.13 points (95% CI 1.19-13.07; P = 0.019), while EQ-5D-5L and PHQ-4 showed no significant change. Satisfaction was high (mean 8.72 ± 1.81). The intervention group had fewer all-cause death or HF hospitalization events (HR 0.39; 95% CI 0.16-0.96; P = 0.041) and lower cardiovascular mortality (1.2% vs. 10.0%; P = 0.035), with a non-significant trend towards fewer HF hospitalizations (sHR 0.43; 95% CI 0.16-1.13; P = 0.087).

CAIS follow-up was feasible, well-received, and low-resource, with exploratory signals of improved outcomes. This pragmatic, scalable approach may enhance HF care and warrants validation in randomized trials.