A protocol for VOICE-HF (voice characteristics evaluation in heart failure)

Abstract

Rationale: In addition to pharmacological treatments, devices and cardiac rehabilitation, Remote Patient Management (RPM) is an important component of heart failure (HF) care. Traditional RPM parameters typically only change shortly before hospitalization and have low sensitivity. A novel, simple, and low-cost approach for early detection of HF decompensation may lie in voice monitoring. To evaluate the potential of voice analysis as a non-invasive monitoring method, the VOICE-HF study is designed to investigate the differences in voice characteristics between acute decompensated heart failure (ADHF), stable CHF, and healthy individuals.

Study population: Patients eligible for this study are divided into three groups. The first group consists of patients hospitalized for ADHF and undergoing intravenous (IV) diuretic treatment. These patients are included in the first 24 hours of admission. The second group consists of stable CHF patients seen at the outpatient clinic. Stable CHF patients are defined by no hospital admission, IV diuretic use, or clinical deterioration in the last four weeks. The third study group consist of healthy individuals, without a personal history of cardiovascular disease or cardiac symptoms at time of inclusion.

Methods

Eligible participants meeting predefined criteria will be enrolled into three cohorts (n=90 per cohort): ADHF, stable CHF, and healthy controls. Participants record 10 predefined sentences and 5 vowels using the ListenHF app. ADHF patients record a second sample on the day of discharge. Stable CHF patients are recruited during outpatient visits and complete a single recording. Healthy volunteers follow the same procedure. Publicly available speech processing models will be used for the analysis of the voice recordings. The model converts raw audio into structured acoustic representations (e.g., log-Mel spectrograms or high-dimensional embeddings), which summarize relevant aspects of the speech signal. These representations will be used in further analyses to explore potential correlations with HF status and differences in the cohorts. Importantly, these models focus only on acoustic features and do not interpret or analyze the linguistic content of speech.

Main study parameters/endpoints: The primary endpoint is the difference in voice characteristics between patients with ADHF, stable CHF and healthy individuals. The secondary endpoint is determination of specific characteristics (e.g shimmer, frequency, pause ratio) related to (AD)HF.

Conclusion

This study, preformed in a large and well-described cohort of ADHF, CHF and healthy controls, represents the first investigation with substantial sample sizes across these groups to explore characteristics associated with (AD)HF. It will provide strong evidence on the potential of voice-monitoring in heart failure and a basis for a follow-up study to assess the predictive potential of voice analysis in detecting ADHF.FlowchartFor image description, please refer to the figure legend and surrounding text.