Advanced Machine Learning Tools to Monitor Biomarkers of Dysphagia: A Wearable Sensor Proof-of-Concept Study

Area of Science:

• Biomedical Engineering
• Neurology
• Machine Learning

Background:

• Difficulty swallowing (dysphagia) is common in neurological disorders, leading to risks like aspiration and malnutrition.
• Current diagnostic methods (imaging, bedside exams) are invasive, costly, or subjective.
• Wearable sensors offer a noninvasive approach to objectively measure swallowing physiology.

Purpose of the Study:

• To develop and validate two novel digital monitoring tools for evaluating swallowing using wearable sensor data and machine learning.
• To create sensitive, clinically meaningful metrics for assessing swallowing performance and impairment severity.

Main Methods:

• Compared biometric swallowing and respiration signals from wearable sensors in post-stroke dysphagia patients and controls.
• Developed two machine learning models: one for classifying swallow impairment severity and another for measuring swallow similarity to normal performance.
• Trained models using kinematic and respiratory features from 505 swallows.

Main Results:

• The developed models provide sensitive, per-swallow metrics, capturing intrasubject variability and patient-specific changes missed by traditional assessments.
• Respiratory-swallow coordination emerged as a key feature for detecting dysphagia and its severity.
• Puree swallows showed significantly greater differences between patients and controls compared to saliva or liquid sips.

Conclusions:

• Interpretable, sensor-based tools are crucial for clinical adoption in dysphagia management.
• These proof-of-concept models offer objective evidence for tracking dysphagia recovery.
• Further validation could enable automated, real-world monitoring of swallowing function across the impairment spectrum.