Voice monitoring to predict heart failure decompensation using AI-based analysis: insights from the Pre-DETECT HF Study

Background Heart failure (HF) is a prevalent chronic condition with a high risk of decompensation. Current remote monitoring strategies largely rely on symptoms and basic vital signs, which often indicate deterioration only after decompensation has occurred, thereby limiting their ability to prevent HF-related hospitalizations. Recent evidence suggests that pulmonary congestion affects vocal tract physiology and leads to subtle yet measurable acoustic alterations. Voice analysis may therefore enable early, non-invasive detection of HF deterioration before clinical symptoms manifest. Purpose To determine whether longitudinal AI-based voice analysis can serve as an early, non-invasive signal of worsening HF (WHF). Methods Pre-DETECT HF is a multicentre, prospective observational study enrolling adult patients hospitalized for acute decompensated HF across three centres in Spain and the Netherlands, with recordings performed in three different languages. Participants completed daily standardized voice recordings and structured health questionnaires during a 6-month follow-up period after discharge. Voice data were transmitted to a secure cloud platform, where acoustic features were extracted and used to train a gradient-boosted classifier. Model performance was assessed using leave-one-patient-out cross-validation, in which each patient’s data were held out during training and subsequently scored using a model trained on all remaining patients. For this interim analysis, predictive performance was evaluated against adjudicated WHF events, defined as HF-related hospitalization or unplanned HF outpatient visits requiring intensification of diuretic therapy. A risk score was generated for each recording, and alerts were triggered when this score exceeded a predefined population-level threshold. A WHF event was considered detected if an alert was active within the 30-day pre-event lead time window. Alerts that did not precede any cardiovascular event were classified as unexplained. Results At interim analysis, 114 participants (mean age 69 ± 12 years; 65 male) contributed 12,361 analysed patient-days. HF phenotype distribution was 43% HFrEF, 12% HFmrEF, and 45% HFpEF. Sixty-eight participants were from the Netherlands and 46 from Spain. During the follow-up, 26 WHF events occurred in 12 patients. The model detected 21 events, yielding a sensitivity of 81%. Median effective lead time was 29 days (IQR 19–52). The unexplained alert rate (UAR) was 2.9 per patient-year. Conclusions Longitudinal AI-based voice analysis was able to identify impending WHF with a low unexplained alert rate and clinically meaningful lead time. These findings support voice analysis as a promising non-invasive digital biomarker for early identification of WHF. Follow-up is ongoing and not all study data are yet available. Final results from the complete dataset will further refine performance estimates and will be presented during the congress.