Abstract
Background
Heart failure (HF) is a common chronic disease that contains the risk decompensation. Most remote HF monitoring relies on symptoms and basic vitals, which signal trouble only when having symptoms of decompensation, failing to prevent hospitalization in half of the cases. Recent evidence suggests that voice alterations may predict clinical worsening, offering a novel and non-invasive biomarker.
Objective
This prospective study aims to validate AI-based voice analysis software for detecting heart failure deterioration after hospital discharge. Primary objective is to estimate the software’s sensitivity in predicting HF events. Secondary objectives include assessing patient adherence and usability of the app, and exploring associations between voice features and blood biomarkers.
Methods
We are conducting a multicentre, prospective observational study including 123 patients hospitalized for acute decompensated heart failure. Patients are enrolled within three days of admission to the cardiology ward, irrespective of HF subtype. As of mid-June 2025, 60 patients have been enrolled, yielding over 14,000 voice recordings. During hospitalisation, participants record daily voice samples while clinically volume overloaded. The final recording prior to discharge, once recompensation is achieved, serves as the individualised baseline (=reference) of the voice pattern. Participants continue daily voice recordings and structured health questionnaires after discharge for a six-month follow-up period, with evaluations at months 1, 3 and 6. Voice recordings consist of 3 vowels and 2 pre-defined sentences. All data are transmitted to a secure cloud infrastructure, where AI-based voice analysis is performed. If symptom thresholds are exceeded, health care professionals (HCP) receive alerts about the potential risk for HF-related decompensation. HCPs may adjust therapy upon their own discretion. Voice-based predictions will be analysed retrospectively. Outcomes are based on HF deterioration events, which was defined as mortality, HF related hospitalisation and intensifying HF therapy due to worsening of HF. Additionally, blood samples collected at baseline, 3, and 6 months will be used to explore the relationship between voice characteristics and conventional or novel blood biomarkers.
Conclusions
relevance The Pre-DETECT HF study explores whether a vocal biomarker could be used as a predictor for HF deterioration. If successful, this approach could enhance remote heart failure monitoring and allow for more timely intervention in patients with HF.