🤖 machine learning

Automated detection of pediatric congenital heart disease from phonocardiograms using deep and handcrafted feature fusion

This paper presents a deep and handcrafted feature fusion model that analyzes phonocardiogram recordings from 751 pediatric patients to achieve 92% accuracy in the automated, cost-effective detection of congenital heart disease, offering a promising screening tool for low-resource settings.

Original authors: Abdul Jabbar, Ethan Grooby, Yang Yi Poh, Khawza I. Ahmad, Md Hassanuzzaman, Raqibul Mostafa, Ahsan H. Khandoker, Faezeh Marzbanrad

Published 2026-04-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Abdul Jabbar, Ethan Grooby, Yang Yi Poh, Khawza I. Ahmad, Md Hassanuzzaman, Raqibul Mostafa, Ahsan H. Khandoker, Faezeh Marzbanrad

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 Problem: Finding Heart Defects in a Noisy Room

Imagine trying to hear a specific bird chirping in a stormy forest. That is what doctors face when listening to a child's heart.

Congenital Heart Disease (CHD) is a birth defect where the heart's structure is slightly "broken" or misshapen. It affects about 1 in every 100 babies. The "gold standard" for finding these defects is an echocardiogram (an ultrasound of the heart), but that machine is expensive, heavy, and requires a highly trained expert to operate. In many parts of the world, especially poorer regions, these resources are scarce.

Doctors often rely on a stethoscope to listen for "murmurs" (abnormal whooshing sounds) that indicate a problem. However, this is like trying to find a needle in a haystack while wearing thick gloves. Children cry, cough, and move, creating background noise that drowns out the subtle heart sounds. Furthermore, not every doctor has the same "ear" for these sounds; one might hear a problem while another misses it.

The Solution: A "Super-Listener" Computer

The authors of this paper built a computer program (an AI) that acts like a super-listening assistant. Its job is to take the messy, noisy recordings from a digital stethoscope and decide: "Is this heart healthy, or does it have a defect?"

They didn't just build one type of listener; they built a hybrid team that combines two different ways of "hearing":

The "Hand-Crafted" Detective (Traditional Features):
Think of this as an old-school music teacher who knows exactly what to look for. The computer was programmed to manually measure specific things about the sound, such as:
- Heart Rate Variability: How much the time between heartbeats changes (like a drummer speeding up or slowing down).
- Spectral Centroid: How "bright" or "sharp" the sound is.
- Spectral Roll-off: How much high-pitched energy is in the sound.
- MFCCs (Mel-Frequency Cepstral Coefficients): This is a fancy way of saying the computer translates the sound into a map that mimics how human ears hear pitch and tone.
The "Deep Learning" Artist (AI Features):
Think of this as a talented artist who has looked at millions of heart sound pictures and learned to spot patterns the human eye (or ear) might miss. They used a 1D Convolutional Neural Network (CNN). Imagine this as a set of filters that scan the sound wave, looking for complex shapes and patterns that don't fit the "healthy" mold.

How They Put It Together: The Fusion

The magic of this paper is Feature Fusion. Instead of letting the "Detective" and the "Artist" work separately, the researchers made them work together.

The computer takes the specific measurements from the Detective.
It combines them with the complex pattern recognition from the Artist.
They are glued together into one big "report card" (a feature vector) that is fed into a final decision-maker.

This is like having a team where one person checks the math and the other checks the art, and then they vote together on the final answer. This approach helps the computer handle the noisy, messy recordings better than using just one method.

The Data: Listening to 751 Kids

To train this team, the researchers went to hospitals in Bangladesh. They recorded heart sounds from 751 children (ranging from 1 month to 16 years old).

They used a special digital stethoscope (the EKO DUO) that is FDA-cleared.
They listened at four different spots on the chest (like checking four different windows in a house): the Mitral, Aortic, Pulmonary, and Tricuspid valves.
The Challenge: The data was "noisy" (kids crying, moving) and "imbalanced" (there were more sick kids in the dataset than healthy ones, which can trick a computer). The researchers used special math tricks to ensure the computer didn't just guess "sick" every time to get a high score.

The Results: A High-Scoring Team

The researchers tested their model on a group of children it had never seen before. Here is how they performed:

Accuracy: 92%. (Out of 100 kids, it got 92 right).
Sensitivity: 91%. (It successfully found 91 out of 100 kids who actually had heart defects. This is crucial so no sick child is missed).
Specificity: 91%. (It correctly identified 91 out of 100 healthy kids as healthy. This is crucial so healthy kids aren't sent to the hospital unnecessarily).
The "Scorecard" (AUROC): 96%. This is a summary score showing the model is excellent at distinguishing between the two groups.

Why This Matters (According to the Paper)

The paper claims this method is a cost-effective screening tool. Because it uses a digital stethoscope (which is cheap and portable) and runs on standard computers, it could be used in places where big ultrasound machines don't exist.

The authors also found that averaging the results from the four different listening spots on the chest worked best. If they just listened to one spot and said "yes" if any spot sounded bad, they made too many mistakes (false alarms). If they required all spots to sound bad, they missed too many cases. The "average" approach found the perfect balance.

What They Didn't Claim

They did not claim this replaces the doctor or the ultrasound machine. It is a screening tool to help decide who needs further testing.
They did not claim it works perfectly on every single type of heart defect (they only distinguished between "CHD" and "Non-CHD," not specific types like "hole in the heart" vs. "leaky valve").
They noted a limitation: The model was trained on data from one country (Bangladesh) with one specific device. They suggest future work should test it on data from other countries and devices to see if it still works.

In short: The researchers built a smart computer program that combines old-school sound measurements with modern AI pattern recognition. It listens to children's hearts through a digital stethoscope and can tell, with about 92% accuracy, if a child likely has a heart defect, offering a potential lifeline for areas where advanced medical equipment is hard to find.

1. Problem Statement

Congenital Heart Defects (CHD) are the most common birth defect, affecting approximately 1% of all live births worldwide. While echocardiography is the gold standard for diagnosis, it is costly, resource-intensive, and inaccessible in low- and middle-income countries (LMICs).

Clinical Challenges: Diagnosis relies heavily on auscultation (listening to heart sounds), which requires high levels of expertise. However, pediatric heart sounds are often faint, contaminated by noise (breathing, crying), and differ significantly from adult patterns.
Limitations of Current AI: Existing deep learning models often rely on raw signal segmentation (which is subjective) or lack the interpretability of clinical features. Many studies focus on adult populations or noise detection rather than specific CHD diagnosis, leading to performance gaps in pediatric screening within resource-constrained environments.

2. Methodology

The authors propose a hybrid deep learning model that fuses handcrafted features (domain-specific) with deep features (data-driven) to classify pediatric patients into CHD and non-CHD categories using phonocardiogram (PCG) signals.

A. Dataset

Source: Collected from the Bangladesh Shishu (Children's) Hospital and the National Heart Foundation Hospital.
Subjects: 751 pediatric patients (aged 1 month to 16 years).
Data Volume: 3,004 audio recordings (15 seconds each) with a sampling rate of 4,000 Hz.
Recording Sites: Four auscultation points: Mitral Valve (MV), Aortic Valve (AV), Pulmonary Valve (PV), and Tricuspid Valve (TV).
Labels: Binary classification (CHD vs. Non-CHD), confirmed by cardiologists via echocardiography, CT, X-ray, and ECG.
Imbalance: The dataset exhibits a 63:37 ratio of CHD to non-CHD cases.

B. Preprocessing

Filtering: A 4th-order Butterworth bandpass filter (25–400 Hz) to remove high-frequency noise.
Normalization: Z-score normalization applied to ensure consistent data scaling.

C. Feature Extraction (The Core Innovation)

The model employs a Multi-Domain Feature Fusion Approach:

Handcrafted Features:
- Heart Rate Variability (HRV): Metrics such as RMSSD, SDNN, and NN50.
- Spectral Features: Spectral centroid, roll-off, and contrast.
- Selection: Features were filtered using the Mann-Whitney U-test ( $p < 0.05$ ) to retain only statistically significant ones.
Deep Features (1D CNN):
- Input: Mel-Frequency Cepstral Coefficients (MFCCs), extracted from PCG signals (including windowing, STFT, Mel-filtering, and DCT).
- Architecture: A custom 1D Convolutional Neural Network with a Multi-Path Design.
  - Input MFCCs pass through an initial convolutional layer.
  - The output splits into three parallel convolutional branches to simultaneously learn diverse feature representations.
  - The outputs are concatenated to form a rich deep feature vector.
Fusion: The handcrafted feature vector and the deep feature vector are concatenated into a single feature array.

D. Classification and Training

Classifier: A fully connected (dense) layer with a Softmax activation function.
Loss Function: Binary Cross-Entropy (BCE) with class weighting to address the 63:37 class imbalance.
Decision Aggregation: Since multiple recordings exist per patient (from different valves), the study compared three aggregation methods:
1. At least one positive (Greedy).
2. Majority vote.
3. Average probability (Averaging probabilities across all recordings).
- Result: The Average Probability method was selected as optimal, as it balanced sensitivity and specificity.
Data Split: Patient-wise split (70% training, 20% validation, 10% test) to prevent data leakage.

3. Main Contributions

Hybrid Feature Fusion: Successful integration of clinically interpretable handcrafted features (HRV, spectral) with high-level deep features (1D CNN on MFCCs), leveraging the strengths of both manual engineering and deep learning.
Pediatric Focus: Addressing the unique challenges of pediatric CHD (noise, age variations) using a dataset specifically collected from infants to adolescents in a real-world LMIC environment.
Computational Efficiency: The model avoids heavy architectures like Transformers and achieves high performance with a lightweight 1D CNN suitable for deployment on portable devices (e.g., digital stethoscopes).
Robust Validation: Utilization of patient-wise splitting and 5-fold cross-validation to ensure the model generalizes to new patients rather than just new signal segments.

4. Results

The model demonstrated superior performance compared to state-of-the-art methods (including the winners of the PhysioNet 2022 competition):

Performance on Test Set (Patient-wise):
- Accuracy: 92%
- Sensitivity: 91%
- Specificity: 91%
- AUROC: 96%
- F1-Score: 92%
Cross-Validation: Mean accuracy of 91.0% ± 2.74 over 5 folds, confirming robustness.
Comparison:
- Outperformed Xu et al. (2022) and Lu et al. (2022) in Specificity (92% vs. 65.5% and 75.5%, respectively), which is critical for reducing false positives in screening.
- Achieved comparable accuracy to Hassanuzzaman et al. (2023), but with better specificity and lower computational costs.
Feature Contribution: Analysis showed that MFCC features contributed most to detection, while handcrafted features provided a necessary boost of approximately 3%, validating the fusion strategy.

5. Significance and Impact

Clinical Utility: The high specificity (91%) represents a major advancement, minimizing unnecessary referrals and reducing the burden on limited healthcare resources in LMICs.
Accessibility: By utilizing digital stethoscopes and a lightweight algorithm, this approach offers a cost-effective, scalable screening tool for remote and underserved areas.
Real-Time Potential: The model processes a 15-second sample in approximately 440 ms on a standard CPU, making real-time diagnosis at the point of care feasible.
Future Directions: The authors suggest that future work could include multi-class classification (specific CHD subtypes), integration of ECG data, and multi-center validation to improve generalizability across different devices and geographies.

In summary, this paper presents a highly effective, interpretable, and computationally efficient solution for the early detection of pediatric CHD, bridging the gap between advanced AI and practical clinical application in resource-limited environments.