Differential Network-Based Causal Graph Learning for Cardiovascular Recurrence Risk Prediction and Factor Discovery

This paper proposes the Causal Factor-aware Graph Neural Network (CFGNN), a novel method that utilizes differential networks to model individual-specific factor interactions for accurately predicting myocardial infarction recurrence risk and identifying key causal risk factors from a systemic perspective.

The Gist

Imagine your body is a massive, bustling city. The heart is the central power plant, and the various health factors (like blood pressure, cholesterol, age, or diabetes) are the different departments, traffic lights, and power lines keeping the city running.

When a patient has a heart attack (a "Myocardial Infarction"), it's like a major blackout in that city. The big question doctors face is: "Will the power go out again?"

Currently, doctors try to predict this by looking at a list of symptoms, kind of like checking a checklist. But this paper argues that a simple checklist isn't enough because it misses how these departments talk to each other.

Here is the paper's solution, broken down into simple concepts:

1. The "Differential Network": The Personalized City Map

Most medical models treat everyone the same. They say, "High blood pressure is bad for everyone." But this paper argues that for you specifically, high blood pressure might be the main problem, while for your neighbor, it might be their diet or stress.

The researchers built a "Differential Network."

• The Analogy: Imagine two maps of the same city. One is the "Standard Map" (what usually happens for most people). The other is a "Live Map" of a specific patient.
• The Magic: They subtract the Standard Map from the Live Map. The result? The "Differential Network." This highlights exactly what is different and broken in this specific patient's city. It shows the unique, weird connections between their health factors that no one else has.

2. GraphSMOTE: Filling in the Missing Puzzle Pieces

In medical data, there are usually way more people who don't have a recurrence (the "safe" group) than people who do (the "at-risk" group). It's like trying to learn how to drive a car by only looking at 100 photos of cars that crashed, but you have 1,000 photos of cars that drove safely. The computer gets confused and thinks "safe" is the only answer.

• The Solution: They used a trick called GraphSMOTE.
• The Analogy: Imagine you have a few rare, broken puzzle pieces (the "at-risk" patients). Instead of just ignoring them, the computer creates new, synthetic puzzle pieces that look exactly like the rare ones but are slightly different variations. This fills out the picture so the computer can learn what a "broken" city actually looks like without getting biased.

3. CFGNN: The "Detective" AI

This is the main star of the show: Causal Factor-aware Graph Neural Network (CFGNN).

• The Problem: Sometimes, a computer sees that "People who wear red shirts get heart attacks" and thinks the red shirt caused the heart attack. But really, maybe they wore red shirts because they were at a specific party where the food was bad. The red shirt is just a coincidence (a "trivial" factor).
• The Solution: The CFGNN acts like a super-smart detective. It splits the patient's data into two piles:
  1. The "Real Culprits" (Causal Factors): These are the actual things causing the risk (like a clogged artery).
  2. The "Red Herring" (Trivial Factors): These are things that just happen to be there but don't cause the problem.
• How it works: The AI is trained to ignore the "Red Herring" pile and focus entirely on the "Real Culprits." It asks, "If I change this specific factor, does the risk actually go down?" If yes, it's a key factor. If no, it's ignored.

4. The Results: What Did They Find?

When they tested this on real patients from hospitals, two big things happened:

1. Better Prediction: The AI was much better at guessing who would have another heart attack compared to old-school methods. It was like upgrading from a crystal ball to a weather radar.
2. New Discoveries: They found that while we already knew age and diabetes were risks, the complexity of the heart lesion (how twisted, blocked, or calcified the specific damaged area was) was actually the most important factor.
   • The Analogy: It's not just that the road is blocked; it's how it's blocked. Is it a simple rock? Or is it a tangled mess of vines and boulders? The "messiness" of the blockage matters more than we thought.

5. Men vs. Women: Different Cities, Different Problems

The study also noticed that the "cities" of men and women have different problems:

• Women: Their risks were often linked to "micro" issues—tiny blood vessels, metabolism, and how the body reacted after surgery.
• Men: Their risks were linked to "macro" issues—big blockages, heavy plaque buildup, and smoking.

The Bottom Line

This paper is like giving doctors a personalized GPS for every heart attack patient. Instead of using a generic map for everyone, it builds a custom map showing exactly how that specific person's health factors are interacting. It filters out the noise, finds the real causes, and helps doctors say, "For this patient, we need to fix this specific part of the city to prevent the next blackout."

This leads to better, more targeted treatments that could save lives and improve quality of life for heart patients.

Technical Summary

1. Problem Statement

Cardiovascular disease (CVD), particularly Myocardial Infarction (MI), poses a significant global health burden. While mortality rates have decreased due to treatment advancements, MI patients face a high risk of recurrent cardiovascular events (e.g., heart failure, recurrent MI, arrhythmias).
Key Challenges:

• Prediction Accuracy: Existing methods (traditional statistics and standard machine learning) struggle with high-dimensional clinical data, often failing to capture complex, non-linear interactions among risk factors.
• Interpretability: Current models often act as "black boxes," failing to identify causal risk factors, which limits their utility in clinical decision-making.
• Data Imbalance: Medical datasets often suffer from class imbalance (fewer recurrence cases than non-recurrence cases), leading to biased models.
• Individual Heterogeneity: Risk factor interactions vary significantly between individuals, which group-level statistical models often overlook.

2. Methodology

The authors propose a three-stage pipeline: Differential Network Construction, GraphSMOTE, and the Causal Factor-aware Graph Neural Network (CFGNN).

A. Individual-Level Differential Network Construction

Instead of using a static global graph, the authors construct a unique differential network for each patient to capture individual-specific deviations in factor relationships.

• Reference Set: A set of reference samples ($P_{ref}$) consisting of patients who did not experience recurrence is used to establish a baseline correlation matrix ($PCC_{Pref}$) between clinical factors.
• Differential Calculation: For a new patient $p_l$, the correlation matrix is recalculated including $p_l$ ($PCC_{Pref \cup \{p_l\}}$). The difference matrix $A_l = PCC_{Pref \cup \{p_l\}} - PCC_{Pref}$ represents the unique deviation in factor relationships for that specific patient.
• Graph Formation: Each patient is represented as a graph where nodes are clinical factors and edge weights are defined by these correlation differences.

B. GraphSMOTE (Data Augmentation)

To address class imbalance in the graph-structured data:

• The method extends the traditional SMOTE algorithm to the graph domain.
• It calculates the distance between adjacency matrices of minority-class graphs.
• It performs linear interpolation between a minority sample and its $K$-nearest neighbors to synthesize new, realistic graph samples, preserving structural properties while balancing the dataset.

C. Causal Factor-aware Graph Neural Network (CFGNN)

The core model is designed to distinguish between causal features (directly influencing recurrence) and trivial features (noise or spurious correlations).

• Decomposition Module: Uses a GNN to generate a mask matrix that splits the input graph into a Causal Subgraph ($G_c$) and a Trivial Subgraph ($G_t$).
• Prediction Module:
  • $G_c$ is used for the primary classification task.
  • $G_t$ is used to ensure trivial features do not influence the prediction.
• Loss Functions:
  • Causal Loss ($L_C$): Maximizes the predictive power of the causal subgraph.
  • Trivial Loss ($L_T$): Minimizes the predictive power of the trivial subgraph (forcing it to predict a uniform distribution).
  • HSIC Loss ($L_{HSIC}$): Measures independence between $G_c$ and $G_t$ to ensure they capture distinct information.
  • Causal Invariance Loss ($L_I$): Inspired by causal invariance principles, this ensures the causal subgraph remains stable across different environments (patient populations) by training on mixed representations of $G_c$ and $G_t$.
• Total Loss: $L = L_C + \alpha L_T + \beta L_{HSIC} + L_I$.

3. Key Contributions

1. High-Quality Real-World Dataset: Construction of a comprehensive, multi-center dataset (MID-I and MID-II) containing treatment and follow-up data for MI patients.
2. Differential Network Approach: A novel method to model individual-level factor interactions rather than relying on population-averaged correlations.
3. GraphSMOTE: Extension of SMOTE to graph-structured data to effectively handle class imbalance in medical networks.
4. CFGNN Model: A causal-aware architecture that explicitly separates causal and trivial factors, improving both prediction accuracy and model interpretability.
5. Discovery of Key Factors: Identification of specific, clinically relevant risk factors (e.g., lesion complexity) that traditional methods might miss.

4. Experimental Results

The model was evaluated on two datasets (MID-I: 1,649 patients; MID-II: 955 patients) using 10-fold cross-validation.

• Performance Metrics: CFGNN outperformed all baselines (Logistic Regression, SVM, KNN, MLP, KAN, GCN, GAT, GIN, ERM, DIR, etc.) across Accuracy, Precision, Recall, F1-Score, and AUC.
  • MID-I: CFGNN achieved 90.35% Accuracy and 90.64% F1-Score, significantly outperforming the next best method (DIR at 88.53% Acc).
  • MID-II: CFGNN achieved 72.58% Accuracy and 63.86% F1-Score, again outperforming all competitors.
• Ablation Studies: Removing any component (Causal Loss, Trivial Loss, HSIC Loss, or Invariance Loss) resulted in performance degradation, confirming the necessity of each module.
• Factor Discovery:
  • The model identified Lesion Complexity (a composite of SYNTAX score, bifurcation, occlusion, calcification, and vessel count) as the most significant risk factor, surpassing traditional factors like age and medical history.
  • Gender-Specific Analysis:
    • Women: Risk factors leaned towards microvascular dysfunction, endothelial impairment, and metabolic disorders.
    • Men: Risk factors were concentrated on macrovascular lesions, plaque burden, and traditional obstructive atherosclerosis.
• Case Studies: Visual analysis of individual differential networks revealed unique topological patterns for recurrence vs. non-recurrence patients, highlighting the model's ability to capture individual heterogeneity.

5. Significance and Impact

• Clinical Decision Support: By identifying causal risk factors rather than just correlations, the model provides actionable insights for clinicians. It suggests that secondary prevention strategies should be tailored based on individual lesion characteristics and gender-specific risk profiles.
• Personalized Medicine: The differential network approach acknowledges that risk factor interactions are patient-specific, moving away from "one-size-fits-all" risk scoring.
• Methodological Advancement: The integration of causal inference with graph neural networks and graph-based data augmentation (GraphSMOTE) offers a robust framework for handling complex, imbalanced, and heterogeneous medical data.
• Targeted Interventions: The discovery that lesion complexity is a dominant predictor suggests that clinicians should prioritize detailed morphological assessments of coronary lesions to better stratify recurrence risk.

In summary, this paper presents a sophisticated deep learning framework that not only predicts cardiovascular recurrence with high accuracy but also uncovers the underlying causal mechanisms, offering a new paradigm for precision medicine in cardiology.