Deep learning approach for dysphagia detection by syllable-based speech analysis with daily conversations

Area of Science:

• Gerontology
• Speech-language pathology
• Artificial intelligence in healthcare

Background:

• Dysphagia (difficulty swallowing) is prevalent in older adults, increasing risks of severe health issues.
• Early detection of dysphagia is crucial for timely intervention and management.
• Existing diagnostic methods may be invasive or not suitable for everyday environments.

Purpose of the Study:

• To evaluate a novel deep learning model for diagnosing dysphagia using syllable-segmented speech data.
• To assess the model's effectiveness in distinguishing between individuals with and without dysphagia.
• To explore the potential of AI for non-invasive, early dysphagia detection in daily life.

Main Methods:

• Collected audio data from 16 dysphagia patients and 24 controls during daily conversations.
• Utilized a speech-to-text model to segment audio into syllables.
• Applied a convolutional neural network for binary classification to identify dysphagia.
• Validated the model using videofluoroscopic swallowing study results.

Main Results:

• The syllable-segmented analysis achieved a diagnostic accuracy of 0.794, sensitivity of 0.901, and specificity of 0.687.
• At the individual level, the model demonstrated an overall accuracy of 0.900 and an AUC of 0.953.
• The deep learning model showed high performance in differentiating dysphagia patients from controls.

Conclusions:

• Deep learning analysis of syllable-segmented speech is a promising tool for early dysphagia detection.
• The developed model offers a non-invasive, simple, and potentially cost-effective method for screening dysphagia.
• This AI-driven approach could facilitate dysphagia diagnosis in everyday settings, improving geriatric care.