Reliable XAI Explanations in Sudden Cardiac Death Prediction for Chagas Cardiomyopathy

This paper proposes a logic-based explainability method with correctness guarantees that achieves 100% explanation fidelity and superior robustness compared to heuristic approaches, thereby enhancing clinical trust and facilitating the deployment of AI-driven tools for predicting sudden cardiac death in Chagas cardiomyopathy patients.

The Gist

Imagine you are a doctor trying to predict which patients with Chagas disease (a heart condition caused by a parasite common in Latin America) might suffer Sudden Cardiac Death (SCD). It's like trying to predict a lightning strike: it happens fast, it's rare, and it's terrifyingly hard to see coming.

For a long time, doctors have used "Black Box" AI models to help with this. Think of these models like a magic 8-ball. You shake it, it gives you an answer ("High Risk" or "Low Risk"), but it refuses to tell you why. If a doctor asks, "Why did you say this patient is at risk?" the magic 8-ball just shrugs. This makes doctors nervous. They can't trust a tool they don't understand, especially when lives are on the line.

The Problem with Current "Explanations"

Some researchers tried to fix this by creating "explanation tools" (like LIME or Anchors). Imagine these tools as guessing games. They look at the magic 8-ball's answer and say, "I think it guessed 'High Risk' because the patient is old." But because they are just guessing, they might be wrong. Sometimes, they might give you the same reason for two completely different patients, or they might miss the real reason entirely. It's like a weather forecaster who says, "It's raining because the clouds are gray," but sometimes it rains when the sky is blue, and their explanation doesn't actually match the reality.

The Solution: The "Logic Detective"

This paper introduces a new approach called Logic-Based Explainability. Instead of a guessing game, think of this new method as a Logic Detective or a Mathematical Proof.

Here is how it works, using a simple analogy:

1. The Training: The researchers built a super-smart AI (using a tool called XGBoost) that is like a master chef. This chef has tasted thousands of patient records and learned exactly which ingredients (medical data points like heart size, rhythm, and blood tests) lead to a "bad meal" (Sudden Death). This chef is incredibly accurate (95% success rate).
2. The Problem: The chef can cook the perfect meal, but if you ask, "Why did you add salt?" the chef just says, "Because I felt like it."
3. The Detective's Job: The new "Logic Detective" steps in. Instead of guessing, it looks at the chef's recipe book (the internal math of the AI) and writes down a strict rule.
   • Example: "If the heart chamber is bigger than 4.5cm AND the blood flow is slow, THEN the risk is high."
4. The Guarantee: The best part? The Detective doesn't just guess. It uses mathematical proof (specifically, a type of logic called "First-Order Logic") to prove that these rules are 100% correct. It guarantees that if these specific conditions are met, the AI will say "High Risk." There is no guessing, no "maybe," and no "black box."

What They Found

The researchers tested this new "Logic Detective" against the old "Guessing Games" (LIME and Anchors) using real patient data from Brazil.

• Accuracy: The Logic Detective was 100% faithful. Every time it gave an explanation, it was perfectly true to the AI's decision. The old guessing games were only about 75-98% faithful.
• Trust: Because the explanations are mathematically proven, a doctor can look at the rule (e.g., "Patient has a specific heart rhythm issue") and say, "Okay, I understand. The AI isn't guessing; it's following a clear rule I can verify."
• Speed: It wasn't the fastest tool, but it was fast enough to be useful in a hospital, and the trade-off was worth it for the certainty it provided.

Why This Matters

In the world of medicine, especially for neglected diseases like Chagas where data is scarce and resources are tight, trust is everything.

If a doctor uses a tool that gives a "black box" answer, they might ignore it. If they use a tool that gives a "guessing" explanation, they might make a mistake because the explanation was wrong. But if they use this Logic-Based AI, they get a tool that is as accurate as a super-computer but as transparent as a clear window.

In short: This paper teaches us how to turn a "magic 8-ball" into a transparent, rule-following partner that doctors can trust with their patients' lives. It proves that we don't have to choose between a smart AI and a clear explanation; we can have both.

Technical Summary

1. Problem Statement

Sudden Cardiac Death (SCD) in patients with Chagas Cardiomyopathy (CC) is a critical, unpredictable clinical challenge, accounting for approximately 45% of deaths in this population. While Machine Learning (ML) and Artificial Intelligence (AI) models have shown promise in risk stratification, their adoption in clinical settings is hindered by a lack of transparency.

• The "Black Box" Issue: Standard ML models are often perceived as opaque.
• Limitations of Current XAI: Common Explainable AI (XAI) methods like LIME and Anchors rely on heuristic approximations. They lack correctness guarantees, meaning they can produce inconsistent explanations (e.g., two different inputs yielding the same explanation) or fail to accurately reflect the model's true decision boundaries. This lack of fidelity creates a barrier to trust among medical professionals.

2. Methodology

The authors propose a framework that combines a high-performance predictive model with logic-based abductive explanations to ensure 100% fidelity and correctness.

A. Dataset and Preprocessing

• Data Source: Clinical data from the Clementino Fraga Filho University Hospital (Federal University of Rio de Janeiro), spanning 1992–2023.
• Cohort: 120 patients (49 original features) after excluding those with severe cardiac impairment already slated for intervention.
  • Classes: 19 patients with Sudden Death (Class 1), 101 without (Class 0).
  • Features: Derived from echocardiograms, ECGs, and 24-hour Holter monitors (30 binary, 12 continuous, 7 categorical).
• Preprocessing: Categorical features were one-hot encoded, resulting in 63 features.
• Feature Selection:
  • Used XGBoost feature importance to filter features.
  • Tested subsets: All 63, 48 (Positive Importance), Top 20, and Top 10.
  • Result: The Top 20 features yielded the best balance of performance and parsimony. Key features included EVTotal, NSVT (mean duration), LVEDV, LA diameter, EF Teicholz, Rassi score, LVESV, TSH, Dyslipidemia, and Amiodarone.

B. Predictive Model

• Algorithm: XGBoost (Gradient Boosted Trees).
• Rationale: XGBoost is robust for tabular data and its tree structure is directly compatible with logic-based explanation methods.
• Hyperparameter Tuning: Used Optuna with 5-fold stratified cross-validation to maximize Recall (critical for minimizing false negatives in life-threatening conditions).
• Performance: The final model achieved 95.00% Recall and 95.00% AUC.

C. Logic-Based Explanation Framework

Instead of heuristic approximations, the authors implemented a formal logic-based approach using First-Order Logic over Linear Real Arithmetic (LRA) and the Z3 SMT solver.

1. Tree Encoding: The trained XGBoost trees were converted into a set of logical formulas ($E$) representing the decision paths.
2. Instance Representation: A specific patient instance ($I$) was represented as a conjunction of feature values.
3. Prediction Logic: The model's prediction ($D$) was defined as a logical formula where the sum of tree outputs determines the class.
4. Abductive Reasoning: To generate an explanation, the system iteratively removes features from $I$ and checks the satisfiability of the formula $I \land E \land \neg D$.
   • If the formula is satisfiable, the removed feature is non-essential (the model can still predict the same class without it).
   • If the formula is unsatisfiable, the feature is essential to the prediction.
   • The result is a minimal, non-redundant set of features that guarantees the prediction.

3. Key Contributions

1. Correctness Guarantees: The primary contribution is the application of logic-based abductive explanations to a high-stakes medical problem, ensuring 100% explanation fidelity. Unlike LIME/Anchors, these explanations are mathematically proven to be consistent with the model's internal logic.
2. Clinical Adaptation: Successfully adapted these theoretical methods to a real-world, small-sample dataset typical of Neglected Tropical Diseases (NTDs), demonstrating practical utility in a resource-constrained environment.
3. Comparative Evaluation: Provided a rigorous empirical comparison against state-of-the-art heuristic methods (LIME and Anchors) regarding fidelity, explanation size, and runtime.
4. Trustworthy AI: Demonstrated a pathway to integrate AI into clinical decision support systems (CDS) by providing transparent, verifiable reasoning that clinicians can trust.

4. Results

The study compared the proposed Logic-Based method against LIME and Anchors on the same XGBoost model.

Metric	Logic-Based (Proposed)	Anchors	LIME
Fidelity	100%	~84.7% (Class 0) / 74.3% (Class 1)	~98.8% (Class 0) / 97.8% (Class 1)
Explanation Size	6.33 (Class 0) / 3.89 (Class 1)	1.99 (Class 0) / 3.58 (Class 1)	Fixed at 10
Runtime	~0.38s (Class 0) / 0.32s (Class 1)	~0.88s	~0.05s (Fastest)

• Fidelity: The logic-based method achieved perfect fidelity, meaning every synthetic sample generated based on the explanation retained the original classification. LIME and Anchors failed to reach 100%, indicating they sometimes misrepresent the model's decision boundary.
• Feature Consistency: The features appearing most frequently in the logic-based explanations (e.g., EVTotal, NSVT, LVEDV) aligned with clinical expectations, though the frequency did not always perfectly match the raw XGBoost "feature importance" scores, highlighting the nuance of logical necessity vs. statistical correlation.
• Trade-off: While LIME was the fastest, the logic-based method offered a superior balance of speed and absolute reliability, which is critical for medical diagnosis.

5. Significance and Conclusion

This work addresses the "trust gap" in AI-driven healthcare. By proving that logic-based explanations can be applied to complex tree-based models with formal guarantees of correctness, the authors provide a robust solution for SCD prediction in Chagas disease.

• Clinical Impact: The high recall (95%) ensures that at-risk patients are not missed, while the 100% faithful explanations allow doctors to verify why a patient is flagged as high-risk, fostering adoption in endemic regions where resources are scarce.
• Future Directions: The authors suggest extending this approach to multimodal data and further validating the explanations through direct expert evaluation to refine clinical guidelines.

In summary, the paper demonstrates that for high-stakes medical applications, sacrificing the speed of heuristic approximations for the mathematical certainty of logic-based explanations is not only feasible but essential for the safe and effective deployment of AI.