MOE-ECG: Multi-Objective Ensemble Fusion for Robust Atrial Fibrillation Detection Using Electrocardiograms

This paper presents MOE-ECG, a multi-objective ensemble fusion framework that utilizes particle swarm optimization and Dempster-Shafer theory to simultaneously optimize predictive performance and model diversity, achieving robust and accurate atrial fibrillation detection across multiple ECG datasets.

The Gist

Imagine your heart is like a drummer in a band. When everything is healthy, the drummer keeps a steady, rhythmic beat. But when a person has Atrial Fibrillation (AFib), the drummer starts playing a chaotic, erratic solo, skipping beats and rushing ahead. Detecting this "chaos" early is crucial because it can lead to strokes or heart failure.

For years, doctors have listened to these heartbeats using an ECG (a machine that records the heart's electrical rhythm). However, with the rise of smartwatches and wearable health devices, we are now drowning in heart data. A doctor can't possibly sit down and listen to thousands of hours of recordings manually. We need a computer to do it, but computers often make mistakes or get confused when the data looks different from what they were trained on.

This paper introduces a new, smarter computer system called MOE-ECG. Here is how it works, explained through simple analogies:

1. The Problem: The "One-Size-Fits-All" Mistake

Most existing computer programs try to detect AFib using a single "expert" or a group of experts who all think exactly the same way.

• The Analogy: Imagine trying to solve a complex puzzle. If you ask 10 people who all look at the puzzle from the exact same angle, they will all make the same mistake. If the puzzle is slightly different (like a different patient's heart data), they will all fail together. This is called a lack of diversity.

2. The Solution: The "All-Star Dream Team"

The authors built a system that doesn't just pick the "smartest" computer model. Instead, it builds a Dream Team of 22 different computer models (some are good at spotting patterns, others are good at spotting irregularities, some are fast, some are careful).

• The Analogy: Think of this like assembling a sports team. You don't just want 11 strikers who only know how to shoot; you need a goalie, defenders, midfielders, and strikers. You need a mix of skills.
• The Magic Ingredient (MOPSO): The system uses a special algorithm (called Multi-Objective Particle Swarm Optimization) to act as the Team Coach. This coach's job is to find the perfect mix of players. The coach has two rules:
  1. The team must win games (high accuracy).
  2. The players must have different playing styles (high diversity).
     The coach constantly tweaks the lineup to find the perfect balance where the team is both smart and varied.

3. The Decision: The "Council of Elders"

Once the Coach picks the best team, how do they vote on whether a patient has AFib?

• The Analogy: Instead of just taking a simple majority vote (which can be swayed by a loud minority), the system uses a method called Dempster-Shafer Theory.
• Imagine a council of elders. If one elder is 90% sure it's AFib, and another is 80% sure, the system weighs their confidence levels carefully. If two elders strongly disagree, the system acknowledges that uncertainty rather than forcing a wrong answer. It combines their "beliefs" to make a final, highly reliable decision.

4. The Results: Finding the "Sweet Spot"

The researchers tested their system using heartbeats recorded in short chunks (like 20, 60, or 120 beats).

• The Finding: They discovered that looking at 60 beats (about one minute of heart rhythm) was the "Goldilocks" zone.
  • 20 beats was too short; the computer was too jittery and made mistakes.
  • 120 beats was too long; it took too much time to get the answer, which isn't good for emergencies.
  • 60 beats was just right. It gave the computer enough information to be sure, but fast enough to be useful in real-time.

Why This Matters

The MOE-ECG system is like a super-smart, diverse, and cautious medical detective.

• It is more accurate than the best single computer model.
• It is more reliable than just averaging the opinions of many computers.
• It is fast enough to work on wearable devices (like smartwatches) to catch AFib before it causes a stroke.

In short, this paper shows that by building a team of diverse AI models and letting them vote carefully, we can detect heart problems much better than before, potentially saving lives by catching these irregular rhythms earlier and more reliably.

Technical Summary

1. Problem Statement

Atrial Fibrillation (AFib) is the most common sustained cardiac arrhythmia, posing significant clinical and economic burdens. While early detection can reduce mortality, the increasing volume of ECG data from wearable and clinical devices makes manual interpretation impractical.

• Limitations of Current Methods: Existing automated solutions often rely on single classifiers or fixed ensembles optimized solely for predictive accuracy. These approaches frequently neglect model diversity, leading to limited robustness and poor generalization across heterogeneous datasets (e.g., different recording devices, patient demographics, and noise levels).
• Core Challenge: There is a need for a framework that simultaneously optimizes predictive performance and classifier diversity to ensure reliable AFib detection from short ECG segments without overfitting to specific dataset characteristics.

2. Methodology: The MOE-ECG Framework

The authors propose MOE-ECG, a novel framework that treats ensemble selection as a bi-objective optimization problem. The workflow consists of four main stages:

A. Data Preprocessing and Segmentation

• Datasets: Three public PhysioNet databases were used:
  • Training: MIT-BIH Atrial Fibrillation Database (AFDB).
  • Validation: Saitama Heart Database Atrial Fibrillation (SHDB-AF).
  • Testing: Long-Term Atrial Fibrillation Database (LTAFDB).
• Signal Processing: Raw ECG signals were filtered to remove baseline wander and noise. R-peaks were detected using the Pan–Tompkins algorithm.
• Segmentation: Signals were segmented into overlapping windows of 20, 60, and 120 consecutive RR-intervals (RRIs) to capture rhythm dynamics at different temporal scales. A segment was labeled "AFib" if >50% of its RRIs were annotated as such.

B. Feature Extraction and Selection

• Extraction: 15 statistical, temporal, and nonlinear features were extracted from RR-intervals (e.g., Mean_RR, SD_RR, RMSSD, Entropy, Poincaré plot ratios).
• Selection: A two-stage strategy reduced redundancy:
  1. Features were ranked by absolute Pearson correlation with the class label.
  2. Features with high mutual correlation ($|r| > 0.85$) were pruned.
• Result: A compact subset of 10 features (including MedianAD, Entropy, CVI, and AVRR) was selected to capture beat-to-beat irregularity and complexity.

C. Multi-Objective Ensemble Selection (MOPSO)

Instead of using a fixed set of models, the authors employed Multi-Objective Particle Swarm Optimization (MOPSO) to select the optimal subset of classifiers from a pool of 22 heterogeneous models (including SVM, Random Forest, XGBoost, LightGBM, TabNet, MLP, VGG, and ResNet).

• Objective 1 (Performance): Maximize the F1-score of the ensemble on the validation set.
• Objective 2 (Diversity): Maximize the pairwise disagreement among selected classifiers.
• Optimization: MOPSO identifies a Pareto front of non-dominated solutions, balancing accuracy and diversity. The final ensemble is selected from this front based on the best validation F1-score.

D. Uncertainty-Aware Fusion (Dempster–Shafer Theory)

To combine the outputs of the selected classifiers, the framework uses Dempster–Shafer Theory (DST) rather than simple averaging.

• DST treats classifier outputs as "mass assignments" (degrees of belief) over hypotheses.
• It explicitly handles uncertainty and conflict between classifiers, providing a more robust probabilistic decision than standard voting or averaging.

3. Key Contributions

1. Bi-Objective Formulation: The study formulates ensemble selection as a multi-objective problem, explicitly optimizing the trade-off between accuracy and diversity, addressing a key gap in traditional accuracy-driven methods.
2. Diversity-Aware Selection: It introduces an automated strategy using MOPSO to select a compact, complementary subset of classifiers from a large heterogeneous pool, avoiding fixed or heuristic ensemble structures.
3. Uncertainty Fusion: The integration of Dempster–Shafer theory allows for robust decision-making in the presence of classifier disagreement, which is critical for medical diagnostics.
4. Generalizability Validation: The framework was rigorously tested across three independent datasets, demonstrating strong generalization capabilities compared to single best classifiers and standard averaging ensembles.

4. Experimental Results

The model was evaluated on the independent LTAFDB test set using 60-beat segments (identified as the optimal window length balancing latency and accuracy).

• Performance Metrics:
  • Accuracy: 86.7% (on test set), 89.85% (average across runs).
  • Sensitivity (Recall): 98.2% (Highest among compared state-of-the-art methods).
  • Precision: 91.14%.
  • F1-Score: 92.64% (average), 81.5% (specific test set result).
  • AUC: ~0.95.
• Statistical Significance: Holm-adjusted Wilcoxon tests confirmed that MOE-ECG significantly outperformed both the "single best classifier" and the "average ensemble of all classifiers" ($p < 0.05$).
• Comparison with State-of-the-Art:
  • MOE-ECG achieved the highest sensitivity (98.2%) compared to deep learning models (CNNs, LSTMs, Transformers) and other ensemble methods tested on the same LTAFDB dataset.
  • While some deep learning models showed slightly higher accuracy, they often suffered from lower sensitivity or required significantly higher computational resources.
• Segment Length Analysis:
  • 20 beats: High variance, lower specificity.
  • 60 beats: Optimal trade-off between detection performance and latency.
  • 120 beats: Performance plateaued; no significant gain over 60 beats, but increased latency.

5. Significance and Conclusion

• Clinical Impact: The MOE-ECG framework offers a robust, accurate, and computationally efficient solution for AFib detection. Its high sensitivity (98.2%) is crucial for minimizing missed diagnoses in clinical and telehealth settings.
• Efficiency: Unlike deep learning models that require massive datasets and GPU acceleration, MOE-ECG relies on a compact set of handcrafted features and lightweight classifiers, making it suitable for real-time deployment on wearable devices and resource-constrained environments.
• Robustness: By explicitly optimizing for diversity and using DST for fusion, the model demonstrates superior generalization across different ECG recording environments, addressing the "domain shift" problem common in medical AI.
• Future Directions: The authors suggest extending the framework to multi-lead ECGs, adaptive windowing, and comparison with foundation models for raw signal processing.

In summary, MOE-ECG represents a significant advancement in cardiac arrhythmia detection by shifting the focus from "accuracy-only" optimization to a balanced, diversity-aware ensemble approach that ensures reliability in real-world clinical applications.