Semi-Supervised Heart Sound Detection Using Spectrogram Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based abnormal heart sound detection models require substantial training datasets covering all heart disease types, which is challenging due to the susceptibility of heart sound signals to background noise, leading to inaccurate detection of abnormal heart sounds.
Innovation Solution
A semi-supervised learning model is generated using the probability distribution of spectrographic properties from normal heart sound signals, transforming them into vector representations, determining latent vectors, and minimizing an optimization function based on logarithmic cross-entropy and Kullback-Leibler divergence to identify abnormal heart sounds without requiring training data related to heart disease types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning models are used for abnormal heart sound detection, then detection accuracy can be improved with sufficient training data, but obtaining accurate training datasets becomes extremely challenging due to background noise susceptibility
Solution Approach 1:
Instead of training the model to detect abnormal heart sounds directly (supervised learning requiring labeled abnormal data), the patent inverts the approach by training the model to recognize normal heart sounds only. The Variational Autoencoder learns the probability distribution of normal heart sound spectrograms, and any deviation from this distribution indicates abnormality. This eliminates the need for labeled abnormal training data while maintaining detection accuracy.
Solution Approach 2:
The patent introduces spectrogram transformation as an intermediary step between the raw heart sound signals and the detection model. By converting time-domain signals to frequency-domain spectrograms, the model can better capture the probabilistic characteristics of normal heart sounds, making the semi-supervised learning approach more effective in distinguishing abnormalities without requiring labeled abnormal examples.
2Reliability
If more training data covering all heart disease types is collected, then model performance improves, but the complexity and difficulty of data collection and preprocessing increases
Solution Approach 1:
The patent applies partial action by collecting and training only on normal heart sound data, rather than attempting to collect comprehensive data covering all possible heart disease types. This semi-supervised approach uses only the readily available normal data (partial action) to build a model that can then detect abnormalities it hasn't explicitly been trained on, significantly reducing data collection and preprocessing complexity.
3Ease of manufacture
If traditional detection algorithms are used, then implementation is simpler, but sensitivity and specificity in identifying abnormal heart sounds are lower compared to semi-supervised learning models
Solution Approach 1:
The patent replaces traditional rule-based or threshold-based detection algorithms with a data-driven semi-supervised learning model using Variational Autoencoders. This substitution of the detection mechanism enables the system to learn complex patterns and probability distributions of normal heart sounds, achieving superior sensitivity and specificity compared to simpler traditional algorithms.
Data Source
AI summary
The disclosure generally relates to methods and systems for identifying presence of abnormal heart sounds from heart sound signals of a subject being monitored. Conventional Artificial intelligence (AI) based abnormal heart sounds detection models with supervised learning requires a substantial amount of accurate training datasets covering all heart disease types for the training, which is quiet challenging. The present methods and systems solve the problem solves the problem of identifying presence of the abnormal heart sounds using an efficient semi-supervised learning model. The semi-supervised learning model is generated based on probability distribution of spectrographic properties obtained from heart sound signals of healthy subjects. A Kullback-Leibler (KL) divergence between a predefined Gaussian distribution and an encoded probability distribution of the semi-supervised learning model is determined as an anomaly score for identifying the abnormal heart sounds.