🤖 AI

Classification of systolic murmurs in heart sounds using multiresolution complex Gabor dictionary and vision transformer

This paper proposes an automatic systolic murmur classification system that utilizes complex orthogonal matching pursuit with a multiresolution complex Gabor dictionary for feature extraction and a vision transformer for classification, achieving 95.96% accuracy on the CirCor DigiScope dataset.

Original authors: Mahmoud Fakhry, Abeer FathAllah Brery

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahmoud Fakhry, Abeer FathAllah Brery

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Listening to the Heart's "Hum"

Imagine your heart is a busy orchestra. When everything is playing in tune, the music is a steady, rhythmic beat. But sometimes, a valve gets a little loose or a pipe gets a little narrow. This causes the blood to rush through with a "whoosh" or a "hiss" instead of a smooth flow. In medical terms, this is called a murmur.

Doctors have been listening to these sounds for centuries with a stethoscope, but it's hard to tell exactly what is wrong just by ear. Is it a "diamond" shape sound? A "plateau"? A "crescendo"? It's like trying to identify a specific instrument in a noisy crowd just by listening.

This paper introduces a new AI detective that can listen to these heart sounds, break them down into tiny puzzle pieces, and figure out exactly what kind of murmur it is with incredible accuracy (about 96%).

Step 1: The "Magic Dictionary" (Feature Extraction)

The first problem the researchers faced is that heart sounds are messy. They change quickly, and they aren't just simple waves. If you try to analyze them with standard tools (like a basic radio tuner), you miss the details.

The Solution: The Multiresolution Gabor Dictionary
Imagine you have a giant library of "sound building blocks." Some blocks are short and sharp (good for catching quick changes), and some are long and smooth (good for catching deep tones).

The Analogy: Think of this as a Swiss Army Knife of sound. Instead of using just one tool, the system has 8 different "sizes" of tools. It uses a mathematical technique called Complex Orthogonal Matching Pursuit (a fancy way of saying "smart matching") to find the perfect combination of these blocks that reconstructs the heart sound.
The Twist: The researchers didn't just look at one sound clip. They looked at multiple clips from the same heart recording. They forced the system to use the same set of building blocks for all of them.
Why this matters: Imagine trying to recognize a friend's voice. If you hear them whisper once and shout once, it sounds different. But if you realize, "Oh, they are using the same vocabulary in both cases," you can identify them better. This "shared dictionary" trick helps the AI ignore the noise and focus on the true shape of the murmur.

Step 2: Turning Sound into a "Map" (Reshaping)

Once the AI has matched the sound to its building blocks, it has a long list of numbers. That's boring and hard for a computer to visualize.

The Solution: The Time-Frequency Map
The researchers took those numbers and reshaped them into 2D pictures (matrices).

The Analogy: Imagine taking a long strip of film and folding it into a grid. Now, instead of a line, you have a map. On this map, the horizontal axis is Time (when the sound happened) and the vertical axis is Frequency (how high or low the pitch was).
Because they used the "8 different sizes" of tools from Step 1, they ended up with 8 different maps for every single heart sound. Some maps are zoomed in on the time (seeing the quick spikes), and others are zoomed in on the frequency (seeing the deep tones).

Step 3: The "Super-Reader" (The Vision Transformer)

Now the AI has 8 different maps. How does it read them?

The Solution: The Vision Transformer (ViT)
Usually, computers use "Convolutional Neural Networks" (CNNs) to look at images. These are great at spotting edges and shapes, but they can get tunnel vision—they might miss the big picture.

The Analogy: Think of a CNN as a person looking at a painting through a small tube. They see a tiny detail very clearly but don't know how it fits with the rest of the painting.
The Vision Transformer is like a person standing back and looking at the entire painting at once. It uses a mechanism called Self-Attention.
How it works: The AI looks at the "Diamond" shape on one map and instantly asks, "Does this connect to the 'Plateau' shape on the other map?" It connects the dots across the whole image, understanding the global context. It's like a detective who doesn't just look at one clue but sees how every clue in the room relates to every other clue.

The Results: A High-Scoring Detective

The researchers tested this system on a massive database of heart sounds (the CirCor DigiScope dataset) containing four types of murmurs:

Diamond: Starts soft, gets loud, then soft again.
Plateau: Stays loud and steady.
Decrescendo: Starts loud and fades out.
Crescendo: Starts soft and gets louder.

The Score:
The system got it right 95.96% of the time.

It was especially good at spotting the "Diamond" and "Plateau" types (over 98% accuracy).
Even for the rare and tricky "Crescendo" type (which had very few examples in the data), it still managed to get it right more than half the time, which is a huge win for AI dealing with rare diseases.

Why This Matters

It's Robust: By using the "shared dictionary" trick, the system handles the fact that heart sounds vary from person to person.
It's Smart: By combining the "local detail" of CNNs with the "big picture" view of Transformers, it sees both the forest and the trees.
The Future: This could lead to a smartphone app or a simple device that a doctor in a remote village could use to instantly diagnose heart valve problems, saving lives by catching issues early.

In a nutshell: The researchers built a super-smart AI that translates messy heart sounds into detailed 8-layer maps, then uses a "super-reader" to look at the whole picture at once, identifying heart problems with near-perfect accuracy.

1. Problem Statement

Cardiovascular diseases (CVDs) are a leading cause of global mortality, necessitating early and accurate diagnostic tools. Heart murmurs, caused by turbulent blood flow, are critical indicators of cardiac abnormalities (e.g., valve stenosis or regurgitation).

Challenge: Systolic murmurs vary significantly in intensity, pitch, and quality (morphology). Traditional time-domain and frequency-domain analysis methods struggle with the non-stationary and complex nature of these signals.
Specific Difficulty: Existing deep learning models (CNNs, RNNs) often fail to capture long-range dependencies or global context effectively. Furthermore, standard time-frequency representations (like STFT) may lack the adaptive resolution needed to distinguish subtle morphological differences between murmur types (Diamond, Plateau, Decrescendo, Crescendo).
Data Imbalance: Clinical datasets often suffer from severe class imbalance, particularly regarding rare murmur types (e.g., Crescendo), making robust supervised learning difficult.

2. Methodology

The authors propose a hybrid system comprising two main modules: a Feature Extraction Module based on sparse signal processing and a Feature Classification Model based on a Vision Transformer (ViT).

A. Feature Extraction: Multiresolution Complex Gabor Dictionary & COMP

Instead of standard transforms, the system uses a Multiresolution Complex Gabor Dictionary combined with Complex Orthogonal Matching Pursuit (COMP).

Dictionary Construction: A redundant dictionary is built using complex Gabor basis functions (atoms) defined by Gaussian windows modulated by sinusoids. The dictionary spans multiple resolutions ( $j=1$ to $8$), creating a trade-off between temporal and spectral resolution (Heisenberg uncertainty principle).
Sparse Projection (COMP): Murmur segments (512 samples) are projected onto this dictionary. Unlike standard OMP, COMP handles complex-valued atoms, preserving phase information (critical for distinguishing turbulence timing and frequency drift).
Joint Modeling: The system processes multiple segments from a single recording simultaneously using a shared dictionary support. This enforces consistency in the selected basis functions across segments, mitigating intra-recording variability.
Reshaping: The resulting sparse complex weight vectors are split by resolution and reshaped into 2D Time-Frequency (TF) feature matrices ( $A_j$ ). These matrices have variable dimensions (e.g., $2 \times 256$ to $256 \times 2$ ) corresponding to different time-frequency resolutions.

B. Feature Classification: Hybrid CNN-Vision Transformer

The classification model processes the multiple variable-resolution TF matrices using a Vision Transformer (ViT) architecture adapted for signal processing.

Patch Tokenization: Each TF matrix is fed into a dedicated branch containing a 2D Convolutional Neural Network (CNN). The CNN acts as a patch tokenizer, extracting local features and reducing dimensionality to a fixed-size embedding vector (32 dimensions).
Embedding & Concatenation: The embeddings from all resolution branches are concatenated to form a single input matrix for the transformer.
Transformer Encoder:
- Multihead Self-Attention: Captures global dependencies and inter-relations between the different resolution features.
- Feedforward Network (FFN): Uses a 1D CNN (kernel size 1) to perform channel-wise transformations.
- Residual Connections & Layer Norm: Ensures stable gradient flow and training convergence.
Output: A fully connected layer followed by a Softmax layer classifies the input into one of the four murmur types.

3. Key Contributions

Novel Feature Extraction: Development of a feature extraction module using Complex Orthogonal Matching Pursuit (COMP) on a Multiresolution Complex Gabor Dictionary. This captures intrinsic time-frequency features while preserving phase coherence, which is lost in real-valued methods.
Joint Segment Modeling: A mechanism to process multiple murmur segments from a single recording using a shared dictionary support. This enforces consistent time-frequency atom selection, effectively reducing variability and overlap issues within a single patient recording.
Hybrid Architecture: Integration of CNNs for local feature extraction (tokenization) and a Vision Transformer for global context analysis. This leverages the strengths of both architectures to handle the hierarchical and long-range dependencies of heart sound signals.
Comprehensive Evaluation: Rigorous testing on the CirCor DigiScope dataset covering four distinct systolic murmur morphologies, addressing the challenge of severe class imbalance.

4. Experimental Results

The system was evaluated on the CirCor DigiScope dataset (4000 Hz sampling rate) using 5-fold cross-validation. The dataset included 4,337 expert-annotated systolic segments across four classes: Diamond, Plateau, Decrescendo, and Crescendo.

Performance Metrics:
- Best Configuration: "Proposed II" (Joint projection of multiple segments).
- Accuracy: 95.96% (Macro-averaged).
- Specificity: 98.43% (Lowest false positive rate).
- F1-Score: 81.87% (Lower F1 is attributed to extreme class imbalance, specifically the scarcity of Crescendo samples, but the model still achieved 59.10% F1 for this rare class).
Comparison: The proposed system significantly outperformed baseline models:
- ResNet with STFT/CWT/WVD features: ~81–87% accuracy.
- PCTMF-Net (Parallel CNN-Transformer): ~83–87% accuracy.
- The proposed method improved accuracy by approximately 10% over the strongest baseline (CWT + ResNet).
Ablation Study: The study confirmed that using multiple attention heads (specifically 32 heads with 32 channels) yielded optimal performance, balancing the capture of diverse discriminative features with model capacity.

5. Significance and Conclusion

Clinical Impact: The system provides a highly accurate, automated tool for diagnosing systolic murmurs, which are the most common and consistently labeled murmurs in clinical practice. High specificity (98.43%) is crucial for minimizing false alarms in screening scenarios.
Methodological Advancement: The paper demonstrates that sparse representation (via Gabor dictionaries) combined with transformer-based global attention is superior to traditional deep learning approaches for non-stationary biomedical signals.
Robustness: The ability to maintain high performance even with severe class imbalance (e.g., correctly identifying rare Crescendo murmurs) highlights the model's robustness and potential for real-world deployment.
Future Directions: The authors plan to extend this framework to diastolic murmurs (which currently lack consistent morphological labels), validate on larger multicenter datasets, and explore real-time point-of-care deployment.

In summary, this research successfully bridges advanced signal processing (sparse coding with complex Gabor atoms) and modern deep learning (Vision Transformers) to solve a critical problem in cardiovascular diagnostics, achieving state-of-the-art results in systolic murmur classification.