Dyspnea Severity Assessment Based on Vocalization Behavior with Deep Learning on the Telephone.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Pulmonology

Background:

• Dyspnea assessment is crucial for respiratory disease management.
• Current methods can be subjective and require in-person visits.
• Remote monitoring solutions are needed to improve patient accessibility and care.

Purpose of the Study:

• To propose and validate a deep learning system for assessing dyspnea using phone-recorded vocalizations.
• To model spontaneous speech patterns during phonetization for objective dyspnea evaluation.
• To develop a prototype for on-line dyspnea estimation via telephone.

Main Methods:

• Utilized deep learning to analyze vocalizations recorded via phone calls (IVR server).
• Engineered and selected time-independent and time-dependent features from controlled phonetizations.
• Employed k-fold cross-validation and score fusion to optimize model generalization and performance.
• Included 104 participants (34 healthy, 70 respiratory patients).

Main Results:

• Achieved 59% accuracy in estimating modified Medical Research Council (mMRC) scores.
• Reported a root mean square error of 0.98 and an area under the ROC curve of 0.97.
• Demonstrated low false positive (6%) and false negative (11%) rates.
• Developed a functional prototype with an ASR-based segmentation scheme.

Conclusions:

• The proposed deep learning system demonstrates feasibility for remote, phone-based dyspnea assessment.
• Vocalization analysis shows potential as an objective and accessible tool for monitoring respiratory conditions.
• The system offers a promising avenue for improving patient management and healthcare accessibility.