HEART SOUND CLASSIFICATION SYSTEM

Abstract

This innovation pertains to a system and approach for categorizing heart sounds through the integration of Convolutional Neural Networks (CNNs) and Transformer models. The system is designed to improve the accuracy of heart sound classification, facilitating the early discovery and diagnosis of cardiac irregularities by harnessing sophisticated signal processing and machine learning methodologies.

Specification

Description:The system architecture is explained in the below sessions.
[015] Heart Sound Acquisition Unit:
• Digital Stethoscope: Captures high-quality heart sounds.
• Amplifier: Ensures adequate signal strength.
• Analog-to-Digital Converter (ADC): Converts analog signals to digital format.
[016] Signal Preprocessing Module:
• Noise Reduction: Filters out background noise using techniques like wavelet denoising or adaptive filtering.
• Segmentation: Identifies and segments individual heartbeats from the continuous recording.
• Normalization: Standardizes the amplitude and duration of heart sounds.
[017] Feature Extraction Module:
• Time Domain: Extracts features such as amplitude, duration, and intervals.
• Frequency Domain: Analyzes spectral features using Fast Fourier Transform (FFT).
• Time-Frequency Domain: Uses techniques like wavelet transform or Short-Time Fourier Transform (STFT).
• Other Methods: Extracts Mel-frequency cepstral coefficients (MFCCs) for perceptual feature analysis.
[018] CNN-Transformer Hybrid Model:
• CNN Component:
 Input Layer: Accepts preprocessed heart sound signals.
 Convolutional Layers: Extract local features using multiple convolutional layers with ReLU activation and pooling.
• Transformer Component:
 Positional Encoding: Adds positional information to the CNN feature maps.
 Encoder Layers: Captures long-range dependencies using multi-head self-attention and feedforward neural networks.
• Fusion Layer:
 Concatenation: Combines CNN and Transformer outputs.
 Dense Layers: Reduces dimensionality and prepares for classification with ReLU activation and dropout.
• Output Layer:
 Softmax Layer: Produces the final classification output with probabilities for each class.

[019] Classification Module:
• Classifies heart sounds into predefined categories such as normal, murmur, and other cardiac anomalies.
[020] User Interface:
• Visual Display: Displays heart sound waveforms, segmented beats, and classification results.
• Audio Playback: Allows users to listen to the heart sounds.
• Diagnostic Suggestions: Provides recommendations based on classification outcomes.
, Claims:We claim
1. A heart sound classification system comprising:
• A heart sound acquisition unit including a digital stethoscope, amplifier, and analog-to-digital converter;
• A signal preprocessing module for noise reduction, segmentation, and normalization of heart sound signals;
• A feature extraction module for extracting time domain, frequency domain, time-frequency domain, and other perceptual features from preprocessed signals;
• A CNN-Transformer hybrid model comprising:
 A CNN component for local feature extraction;
 A Transformer component for capturing long-range dependencies;
 A fusion layer for combining CNN and Transformer outputs and reducing dimensionality;
 An output layer for generating classification results;
• A classification module for classifying heart sounds into predefined categories;
• A user interface for displaying classification results, allowing audio playback, and providing diagnostic suggestions.
2. The system of claim 1, wherein the signal preprocessing module uses wavelet denoising for noise reduction.
3. The system of claim 1, wherein the feature extraction module uses Fast Fourier Transform (FFT) for frequency domain analysis.
4. The system of claim 1, wherein the CNN-Transformer hybrid model includes multiple convolutional layers with ReLU activation and pooling.
5. The system of claim 1, wherein the Transformer component includes multi-head self-attention mechanisms and feedforward neural networks.
6. The system of claim 1, wherein the fusion layer uses concatenation and dense layers with ReLU activation and dropout.
7. The system of claim 1, wherein the user interface includes visual display, audio playback, and diagnostic suggestions based on classification outcomes.

Documents