Voice-based remote monitoring for the early detection of adverse events in chronic heart failure: rationale and design of the TIM-HF3 voice substudy

Remote monitoring reduces mortality and morbidity in chronic heart failure (HF) but typically relies on multiple sensors and daily expert review, limiting scalability. Voice carries acoustic signatures of pulmonary congestion and autonomic dysregulation, physiological changes that precede cardiac decompensation. The Telemedical Interventional Management in Heart Failure III trial (TIM-HF3) included a pre-specified substudy to evaluate voice as a stand-alone digital biomarker for the early detection of adverse HF events.

To determine the diagnostic accuracy and clinical lead time of a voice-based algorithm in predicting HF-related hospitalisation or unplanned therapy escalation within 30 days, as a pre-specified secondary outcome of the TIM-HF3 main trial.

TIM-HF3, conducted at three German centres, enrolled 105 adults with New York Heart Association class II–III HF, left-ventricular ejection fraction (LVEF) ≤ 45 %, or HF with preserved ejection fraction on chronic diuretics, and ≥1 HF hospitalisation in the previous year. Participants recorded sustained vowels, a 30-second scripted passage, and spontaneous speech weekly for up to 18 months via a CE-marked mobile application. Encrypted audio was stored on a central server for retrospective analysis. Routine remote measurements (body weight, blood pressure, single-lead ECG, and symptom score) were collected as standard of care but excluded from model development. A gradient-boosted decision tree model was trained on selected acoustic features to classify recordings as 'wet' if they were obtained ≤30 days before an adjudicated HF hospitalisation, and as 'dry' otherwise. Model output was a continuous wetness score. Leave-patient-out cross-validation ensured subject-independent evaluation. A blinded clinical endpoint committee adjudicated a composite outcome of primary ICD-10 I50 HF hospitalisation, oral or intravenous diuretic intensification, or emergency HF visit. Alerts issued within 30 days before an event were considered true positives.

The substudy included 105 patients, with a total of 25 000 voice recordings and 47 adjudicated HF-related hospitalisations or urgent interventions. Primary analysis will report the model’s sensitivity and the rate of false-positive alerts per patient-year. Secondary metrics include the area under the receiver-operating-characteristic curve and median lead-time before clinical events.

The TIM-HF3 voice sub-study explores whether weekly voice recordings alone may support early identification of adverse HF events. If successful, this approach could contribute to developing scalable and accessible strategies for remote heart failure management.