Deep Learning Swallow Classification via Cervical Vibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for classifying swallowing impairments, such as dysphagia, using cervical auscultation signals face limitations due to small sample sizes, over-fitting, reliance on pre-determined statistical features, and the use of linear classifiers, which can reduce accuracy and complicate the classification process.
Innovation Solution
The implementation of a deep learning classifier, specifically a Deep Belief network with multiple hidden layers, to analyze cervical auscultation vibration data and classify swallows, allowing for the differentiation between healthy and dysphagic swallows through unsupervised learning and higher-order signal feature analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If pre-determined statistical features and linear classifiers are used, then the classification process is simpler, but the classification accuracy is reduced
Solution Approach 1:
The patent transforms the classification approach by changing from linear classifiers to deep belief networks, which can capture non-linear relationships in the data. This parameter change in the classifier type enables higher accuracy while the automated feature learning compensates for the increased complexity through efficiency gains.
Solution Approach 2:
The patent replaces traditional signal processing methods (mechanical/systematic approach) with deep learning algorithms. By substituting manual feature extraction and linear classification with automated deep belief networks, the system achieves superior accuracy while the computational efficiency offset the algorithmic complexity.
2Productivity
If small sample sizes are used for training, then the training process is faster, but the model over-fits and generalizability is reduced
Solution Approach 1:
The patent applies pre-training to the deep belief network before fine-tuning on the specific swallowing dataset. This preliminary action of pre-training on general data enables the model to learn robust features that generalize better, reducing over-fitting even with limited sample sizes while maintaining reasonable training efficiency.
3Measurement precision
If multiple measurements and signals are incorporated, then the classification completeness is improved, but the hardware complexity and system complexity increase
Solution Approach 1:
The patent extracts and focuses on the most critical signal - cervical auscultation vibrations - while using the deep belief network to automatically learn relevant features from this single modality. This extraction approach achieves comprehensive classification without requiring multiple sensors or measurement modalities, thereby reducing hardware complexity while maintaining classification completeness.
Data Source
AI summary
A method of classifying a swallow of a subject includes obtaining vibration data that is based on and indicative of a number of vibrations resulting from the swallow, and using a computer implemented deep learning classifier to classify the swallow based on the vibration data. Also, a system for classifying a swallow of a subject includes a computing device implementing a deep learning classifier. The computing device includes a processor apparatus structured and configured to receive vibration data that is based on and indicative of a number of vibrations resulting from the swallow, and use the deep learning classifier to classify the swallow based on the vibration data. The deep learning classifier may comprise a single layer or a multi-layer Deep Belief network.