An interpretable and explainable neural network to classify sports-related cardiac arrhythmias in professional football athletes

This study presents an interpretable neural network framework that classifies sports-related cardiac arrhythmias in professional footballers by comparing standard and sinc convolution architectures, revealing that sinc models better capture periodic rhythms like normal sinus rhythm while standard models excel at complex morphological patterns, with both validated through physiological feature attribution.

The Gist

Imagine you are a coach for a professional football team. Your players are in peak physical condition, but sometimes their hearts beat in ways that look strange on a standard medical chart (an ECG). The big challenge? Telling the difference between a heart that is super-fit (like a high-performance sports car engine) and a heart that is sick (like an engine with a hidden crack).

This paper is about building a smart computer program (an AI) to help doctors make that distinction, but with a special twist: the program doesn't just give an answer; it explains why it thinks that way.

Here is the story of how they did it, broken down into simple parts:

1. The Problem: The "Athlete's Heart" Mystery

When regular people exercise, their hearts get stronger. But when elite athletes exercise, their hearts adapt in extreme ways. They might beat very slowly (bradycardia) or show weird patterns on an ECG.

• The Risk: Sometimes, these weird patterns are just signs of being a super-athlete. Other times, they are signs of a dangerous heart condition that could lead to sudden death on the field.
• The Gap: Doctors have plenty of data on "normal" sick people, but very little data on "normal" elite athletes. It's like trying to teach a student to recognize a Ferrari by only showing them pictures of Toyotas.

2. The Solution: Two Different "Detectives"

The researchers built two different AI "detectives" to look at heart signals. They trained them on a massive library of general hospital heart data (the "Toyotas") and then tested them on professional football players (the "Ferraris").

• Detective A (The Pattern Matcher): This detective uses a standard approach. It looks at the whole picture and tries to memorize shapes and squiggles. It's good at spotting complex, messy patterns.
• Detective B (The Frequency Tuner): This detective is special. It uses something called a "Sinc Convolution." Think of this like a radio tuner. Instead of looking at the whole messy signal, it forces the AI to focus only on specific "frequencies" or rhythms, ignoring the static. It's designed to understand the physics of the heartbeat, not just the shape.

3. The "Black Box" Problem

Usually, AI is a "black box." You feed it data, and it spits out an answer, but you have no idea how it decided. In medicine, this is dangerous. If an AI says a player is healthy, the doctor needs to know why.

To fix this, the team used a tool called Grad-CAM.

• The Analogy: Imagine the AI is a student taking a test. Grad-CAM is like a highlighter pen. After the student writes the answer, the highlighter shows exactly which words in the question they looked at to get the right answer.
• The Result: The researchers could see exactly which part of the heartbeat the AI was staring at. Did it look at the "P-wave" (the start of the beat)? Did it look at the "T-wave" (the end)?

4. The Results: Who Won?

The two detectives had different strengths, much like two different types of athletes:

• Detective B (The Frequency Tuner/Sinc) was the Rhythm Master.

  • It was excellent at spotting Normal Sinus Rhythm (a healthy, steady beat).
  • Why? Because it was tuned to listen for the specific "song" of a healthy heart. It ignored the noise and focused on the beat.
  • Weakness: It sometimes got confused by complex, messy shapes that didn't fit a perfect rhythm.

• Detective A (The Pattern Matcher/Standard) was the Shape Shifter.

  • It was better at spotting T-wave inversions and Bundle Branch Blocks (complex structural issues).
  • Why? Because these conditions look like specific "shapes" or "scars" on the graph, and the standard detective is great at recognizing visual patterns.

5. The "Oops" Moment: The Zero-Padding Glitch

There was a funny mistake the AI made. Both detectives sometimes got distracted by zero-padding.

• The Analogy: Imagine you are trying to read a book, but someone taped blank white paper to the beginning and end of every page. The AI got so confused that it started thinking the blank paper was the most important part of the story!
• The Lesson: This showed the researchers that the AI was sometimes paying attention to the "glue" holding the data together, rather than the actual heart data. This is a crucial lesson for making AI safer in the future.

6. The Big Takeaway

This study proves that we can use AI trained on general patients to help diagnose elite athletes, but we need to be careful about how we build the AI.

• If you want to check if a heart is beating in a healthy rhythm, use the Frequency Tuner (Sinc).
• If you want to check for complex structural weirdness, use the Pattern Matcher (Standard).
• Most importantly, by using the "highlighter" (Grad-CAM), doctors can trust the AI because they can see exactly what the computer is looking at.

In a nutshell: The researchers built a smart system that doesn't just guess if a footballer's heart is safe; it points to the specific part of the heartbeat that made it safe, helping doctors make better, safer decisions for their players.

Technical Summary

1. Problem Statement

Clinical Context: Sudden Cardiac Death (SCD) risk is 2–3 times higher in athletes than in non-athletes. Differentiating between benign, exercise-induced cardiac adaptations (e.g., sinus bradycardia, incomplete right bundle branch block) and pathological conditions is a critical challenge in sports medicine.
The Gap:

• Data Scarcity: Open-source ECG datasets specific to professional athletes are scarce.
• Black-Box AI: While AI can classify arrhythmias, standard deep learning models lack interpretability, making it difficult for clinicians to trust their decisions, especially when distinguishing physiological adaptations from pathology.
• Domain Shift: Models trained on general-population hospital data often fail to generalize to the unique physiological profiles of elite athletes due to distribution shifts (e.g., age, heart rate, prevalence of specific rhythms).

2. Methodology

The authors propose a novel Explainable AI (xAI) pipeline that integrates four axes: data explainability, model interpretability, post-hoc explainability, and systematic assessment.

A. Datasets & Domain Adaptation

• Source Domain (Training): PhysioNet Challenge 21 (88,253 ECGs from the general population, mean age ~60, 30 arrhythmia classes).
• Target Domain (Testing): PF12RED (161 12-lead ECGs from Spanish La Liga professional footballers, mean age ~26, resting state).
• Strategy: The models were trained on the large, heterogeneous general-population dataset and tested on the scarce athlete dataset to evaluate domain adaptation capabilities.

B. Neural Network Architectures

Two distinct architectures were developed and compared within a Residual Neural Network (ResNet) framework with multi-head attention:

1. Standard Convolution Head: Uses regular convolutional layers that learn arbitrary feature patterns without inherent physiological constraints.
2. Interpretable Sinc Convolution Head: Uses a Sinc convolution layer in the first stage. This layer constrains the network to learn "low- and high-cut-off frequencies" of band-pass filters.
   • Rationale: This forces the network to focus on physiologically relevant frequency ranges (e.g., P-wave: 5–30Hz, QRS: 8–50Hz, T-wave: 0–10Hz) rather than arbitrary patterns, enhancing model interpretability.

C. Training & Evaluation Strategy

• Ensemble: The final model is an ensemble of three subunits using majority voting.
• Optimization: Class-specific thresholds were optimized using differential evolution (a stochastic search algorithm).
• Metrics: Performance was evaluated using AUROC, AUPRC, and F1-score to address severe class imbalance.
• Explainability Tool: Grad-CAM (Gradient-weighted Class Activation Mapping) was used to generate heatmaps highlighting the specific ECG segments (e.g., PR interval, T-wave) driving the classification.

3. Key Contributions

1. Novel xAI Framework: The study moves beyond single-method explainability by integrating data analysis, intrinsic model interpretability (Sinc layers), and post-hoc visualization (Grad-CAM) into a cohesive pipeline.
2. Sinc Convolution for Cardiac AI: Demonstrates that constraining neural networks with learnable band-pass filters (Sinc layers) improves the detection of periodic rhythms (like Normal Sinus Rhythm) by aligning the model's internal logic with cardiac physiology.
3. Domain Adaptation Feasibility: Proves that models trained on general-population data can be adapted to classify arrhythmias in professional athletes, despite significant demographic and physiological differences.
4. Critical Analysis of Artifacts: The study uniquely identifies that zero-padding (a pre-processing technique) is incorrectly highlighted by Grad-CAM as a critical feature, particularly in Sinc models, revealing a vulnerability in how these models handle signal boundaries.

4. Key Results

The models were tested on four specific classes: Normal Sinus Rhythm (NSR), Sinus Bradycardia (SB), Incomplete Right Bundle Branch Block (IRBBB), and T-wave Inversion (TWI).

Metric	Sinc Convolution (Interpretable)	Standard Convolution (Regular)	Interpretation
NSR (AUROC)	0.75	0.70	Sinc better captures periodic rhythms.
SB (AUROC)	0.73	0.74	Comparable performance.
IRBBB (AUROC)	0.58	0.66	Regular conv better captures complex morphology.
TWI (AUROC)	0.54	0.59	Regular conv better captures morphological changes.
F1-Score (SB)	0.81	0.81	Both models excel at detecting athlete bradycardia.
F1-Score (IRBBB)	0.07	0.19	Both struggle due to class imbalance, but regular is better.

Grad-CAM Findings:

• Sinc Model: Focused correctly on physiologically relevant segments (PR interval for NSR/SB; T-wave for TWI). However, it showed high sensitivity to zero-padding, treating it as a critical feature.
• Regular Model: Showed broader activation patterns, sometimes focusing on the T-wave for NSR, and was less sensitive to zero-padding artifacts.

5. Significance and Limitations

Significance:

• Clinical Trust: By demonstrating that Sinc layers focus on known physiological markers (like the PR interval), the study increases the "trustworthiness" of AI in sports cardiology.
• Architectural Guidance: The results suggest a hybrid approach is optimal: use Sinc layers for rhythm-based detection (periodicity) and Standard layers for morphology-based detection (structural abnormalities).
• Scalability: Validates the use of large public datasets (PhysioNet) to train models for niche populations (athletes) when specific data is unavailable.

Limitations:

• Class Imbalance: Severe imbalance in the training data (e.g., IRBBB was only 1.4% in PhysioNet vs. 14.3% in athletes) led to poor performance in detecting IRBBB and TWI.
• Zero-Padding Artifacts: The models, especially Sinc, learned to rely on pre-processing artifacts (zero-padding) rather than just biological signals, indicating a need for better signal processing techniques (e.g., Dynamic Time Warping).
• Risk Stratification: The current loss functions are tuned for the general population, not the specific risk profile of athletes where some "abnormalities" are benign.

Conclusion:
The paper successfully demonstrates that an interpretable neural network using Sinc convolutions can effectively classify sports-related arrhythmias via domain adaptation. While standard convolutional networks are superior for complex morphological patterns, Sinc networks offer superior interpretability and performance for rhythmic conditions, providing a robust foundation for future AI-driven sports cardiology screening tools.