Autonomous Swallow Segment Extraction Using Deep Learning in Neck-Sensor Vibratory Signals From Patients With...

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Dysphagia, a swallowing disorder, increases risks of aspiration pneumonia and mortality.
• Current diagnosis relies on videofluoroscopic swallowing studies (VFSS), which have accessibility and feasibility limitations.
• Acceleration signals from neck sensors offer a potential radiation-free alternative for dysphagia assessment.

Purpose of the Study:

• To develop and evaluate a deep learning model for automatic dysphagia detection using neck acceleration signals.
• To compare the performance of the proposed model against other deep network variants.

Main Methods:

• A hybrid deep convolutional recurrent neural network was employed for multi-level feature extraction from multi-channel swallowing acceleration signals.
• The model was trained and validated on data from 3144 swallows across 248 patients with suspected dysphagia, alongside VFSS images.
• Performance was evaluated using the area under the receiver operating characteristic curve (AUC).

Main Results:

• The developed deep learning network achieved a superior AUC of 0.82 (95% CI: 0.807-0.841) in detecting swallow segments.
• The model demonstrated agreement with up to 90% of gold-standard labeled segments.
• The network effectively performed multi-level feature extraction for accurate swallow segmentation.

Conclusions:

• The proposed hybrid deep learning model accurately detects swallowing segments using neck acceleration signals.
• This method provides a promising, radiation-free alternative for dysphagia diagnosis, overcoming limitations of VFSS.
• Further research can explore integrating this technology into clinical practice for improved dysphagia management.