Abstract:Objective: Dysphagia is a significant ischemic stroke complication that can lead to aspiration. Identification of at-risk patients can be logistically difficult and costly. Researchers investigated whether quantitative ultrasound assessment of hyoid bone movement during induced swallowing would predict failure of videofluoroscopy (VFS) or fiberoptic endoscopic evaluation of swallowing (FEES), as determined by a penetration-aspiration scale (PAS) score. Additionally, ability of a machine learning (ML) algorithm to predict PAS success or failure from real-time ultrasound video recordings was assessed. Methods: A prospective, single-blinded, observational pilot study was conducted from June 2019 through March 2020 at a comprehensive stroke center on a convenience sample of patients admitted with diagnosis of acute ischemic stroke undergoing VFS or FEES as part of dysphagia assessment. Researchers performed a midsagittal airway ultrasound during swallowing in patients receiving an objective swallowing assessment by speech language pathologists who were blinded to ultrasound results. Sonologists measured hyoid bone movement, and researchers then constructed an ML algorithm designed for real-time video analysis using a long short-term memory network with an embedded VGG16 convolutional neural network. Results: Videos from 69 patients were obtained with their respective PAS results. In total, 90% of available videos were used for algorithm training. After training, the ML algorithm was challenged with the 10% previously unseen videos and then compared with PAS outcomes. Statistical analysis included logistic regression and correlation matrix testing on human ultrasound measurements. Cohen's κ was calculated to compare deep learning algorithm prediction with PAS results. Measurement of hyoid bone elevation, forward displacement, total displacement and mandible length was unable to predict PAS assessment outcome (p values = 0.36, 0.13, 0.11 and 0.32, respectively). The ML algorithm showed substantial agreement with PAS testing results for predicting test outcome (κ = 0.79; 95% confidence interval: 0.52-1.0) CONCLUSION: Manual ultrasound measurement of hyoid movement during swallowing in stroke patients failed to predict PAS swallowing results. However, an ML algorithm showed substantial agreement with PAS results despite a small data set, which could greatly improve access to dysphagia assessment in the future.

Deep learning models for bolus segmentation in videofluoroscopic swallow studies

Automated pharyngeal phase detection and bolus localization in videofluoroscopic swallowing study: Killing two birds with one stone?

Automatic Annotation of Cervical Vertebrae in Videofluoroscopy Images Via Deep Learning

Recent Advancements and Future Directions in Automatic Swallowing Analysis Via Videofluoroscopy: A Review

PECI-Net: Bolus segmentation from video fluoroscopic swallowing study images using preprocessing ensemble and cascaded inference

Automatic Tracking of Hyoid Bone Displacement and Rotation Relative to Cervical Vertebrae in Videofluoroscopic Swallow Studies Using Deep Learning

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

The Development of an Artificial Intelligence Video Analysis-Based Web Application to Diagnose Oropharyngeal Dysphagia: A Pilot Study

Machine-learning assisted swallowing assessment: a deep learning-based quality improvement tool to screen for post-stroke dysphagia

A deep-learning-based unsupervised model on esophageal manometry using variational autoencoder

Predicting Swallowing Recovery in Subacute Stroke Patients via Temporal and Spatial Parameters of Videofluoroscopy

Bedside Ultrasound to Identify and Predict Severity of Dysphagia Following Ischemic Stroke: Human Versus Artificial Intelligence

sEMG-based automatic characterization of swallowed materials

Automated Laryngeal Invasion Detector of Boluses in Videofluoroscopic Swallowing Study Videos Using Action Recognition-Based Networks

A Novel Approach for Swallow Detection by Fusing Throat Acceleration and PPG Signal.

A deep learning approach to dysphagia-aspiration detecting algorithm through pre- and post-swallowing voice changes

Video-TransUNet: Temporally Blended Vision Transformer for CT VFSS Instance Segmentation

Upper Esophageal Sphincter Opening Segmentation with Convolutional Recurrent Neural Networks in High Resolution Cervical Auscultation

Multi-Scale Spatial and Temporal Speech Associations to Swallowing for Dysphagia Screening.

Expert-level aspiration and penetration detection during flexible endoscopic evaluation of swallowing with artificial intelligence-assisted diagnosis