Voice features and deep learning models for identifying acute decompensated heart failure.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Acute decompensated heart failure (ADHF) presents as systemic congestion requiring hospitalization.
• Early, non-invasive detection of ADHF decompensation is a significant clinical challenge.
• Voice characteristics may serve as a potential non-invasive biomarker for ADHF.

Purpose of the Study:

• To investigate differences in voice characteristics between ADHF patients during hospitalization and after recovery.
• To assess the feasibility of using voice analysis and deep learning for ADHF detection.

Main Methods:

• Prospective enrollment of 100 ADHF patients requiring hospitalization.
• Exclusion of patients with confounding conditions (e.g., infection, vocal cord disease).
• Voice recordings at admission (ADHF state) and discharge (recovered state).
• Extraction of low-level audio features and Mel-spectrograms for deep learning models.
• Evaluation using area under the receiver operating characteristic curve (AUC).

Main Results:

• Harmonics-to-noise ratio and Shimmer showed potential in distinguishing between decompensated and recovered states.
• The ConvNeXt-large deep learning model achieved a mean AUC of 0.83 at the patient-state level.
• Segment-level analysis yielded higher discriminative performance with AUC up to 0.87.

Conclusions:

• Significant differences in voice characteristics exist between ADHF admission and discharge states.
• Deep learning-based voice analysis can effectively differentiate between acute decompensated and recovered states in ADHF patients.
• Voice analysis offers a promising novel, non-invasive approach for ADHF detection.