Background
Pulmonary congestion is a major driver of heart failure (HF) decompensation predicts hospitalization. Home pulmonary arterial pressure (PAP) monitoring with implantable sensors enables the early detection of congestion and allows for timely therapeutic interventions. However, the invasive nature and high costs of these devices limit their widespread use. Recent advances in voice analysis suggest that fluctuations in PAP may be reflected by subtle changes in voice characteristics, providing a promising non-invasive approach for the remote monitoring of pulmonary congestion.
Objective
This multicenter study investigates whether AI-driven voice analysis can detect changes in PAP in HF patients with an implanted pulmonary artery pressure sensor. By comparing daily voice recordings with sensor-derived PAP values, the study aims to identify vocal biomarkers that indicate clinically relevant hemodynamic changes.
Methods
This is an ongoing multicenter, prospective, observational study enrolling patients with a previously implanted PAP sensor (CardioMEMS™, Abbott, Chicago, IL, USA) who had already been included in a HF telemonitoring program as per clinical indication. Patients from U.S. and German study sites provide daily voice recordings in English or German following a predefined protocol using a smartphone application (Noah Labs App, Noah Labs GmbH, Berlin, Germany). PAP measurements are transmitted shortly before or after each voice recording. Voice recordings are retrospectively analyzed and correlated with PAP trends. Clinical outcomes are prospectively assessed during the study period. The study includes a three-month recruitment phase per site and a six-month follow-up per participant. The primary endpoint is the sensitivity and specificity of AI-based voice models in detecting PAP changes of <5 mmHg, 5–10 mmHg, and >10 mmHg. Secondary endpoints include the feasibility of and compliance with daily voice-based remote monitoring.
Results
As of the current interim analysis, 27 out of 80 patients have been enrolled across three participating sites, contributing a total of 14,096 voice recordings and 2,577 matching PAP measurements. Recruitment, data processing, and model training are ongoing.
Conclusion
The interim analysis results demonstrate high patient adherence and technical feasibility of daily voice and PAP data collection, supporting the potential role of voice as a non-invasive tool for remote pulmonary pressure monitoring in HF patients. If successful, this approach could provide a scalable, cost-effective alternative to implantable pressure sensors, facilitating early detection of hemodynamic deterioration and supporting optimized HF management.
