🤖 machine learning

CW-B: Class Weighted Boosting Framework for Imbalance Resilient Multi Class Cardiac Phenotyping

The paper introduces CW-B, a class-weighted XGBoost framework that effectively addresses real-world data imbalance and missingness to achieve superior performance in five-class cardiac discharge phenotyping while maintaining clinical interpretability.

Original authors: Sijia Li, Xiaoyu Tan, Chen Zhan, Yuanji Ma, Haoyu Wang, Xihe Qiu

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sijia Li, Xiaoyu Tan, Chen Zhan, Yuanji Ma, Haoyu Wang, Xihe Qiu

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 Picture: Sorting Patients into Groups

Imagine a hospital is a busy train station. Every day, thousands of patients arrive with different heart conditions. When they leave (get discharged), the doctors need to sort them into one of five specific groups (phenotypes) to decide their future care plan.

The groups are:

Stable CAD: A patient with a heart condition that is currently calm and predictable.
ACS: A patient having a sudden, dangerous heart emergency (like a heart attack).
Old MI: A patient who had a heart attack in the past.
CAS: A patient suspected of having heart issues, but tests show their arteries aren't blocked enough to be a major crisis yet.
Non-CAD: A patient whose heart issues are actually caused by something else entirely.

The Problem:
In the real world, patient records are messy.

The "Missing" Problem: Sometimes doctors forget to write down a test result, or a machine fails to record a number. It's like a student turning in a test with blank answers.
The "Imbalance" Problem: Some groups are huge (like the "Stable" group), while others are tiny (like the "Old MI" group). If you train a computer to guess the most common group every time, it will get a high score overall, but it will completely miss the rare, dangerous groups. Missing a "dangerous" patient is like missing a bomb in a security line—it's a much worse mistake than missing a harmless one.

The Solution: CW-B (The Smart Sorter)

The authors built a new computer program called CW-B. Think of it as a super-smart sorting machine designed specifically to handle messy, unbalanced data without making dangerous mistakes.

Here is how CW-B works, using three main tricks:

1. The "Missing Note" Flag

When a piece of data is missing (like a blank test score), old programs usually just guess a number to fill the hole. CW-B does something smarter. It adds a red flag next to that number.

Analogy: Imagine you are grading a test. If a student leaves a question blank, you don't just guess they got it right or wrong. You put a sticky note on it that says, "This was left blank." CW-B treats that "blankness" as important information. It tells the computer, "The fact that this was missing might actually tell us something about the patient's condition."

2. The "Fairness Scale" (Class Weighting)

Because some patient groups are rare, the computer naturally ignores them. CW-B puts a heavy weight on the rare groups and a lighter weight on the common groups.

Analogy: Imagine a teacher grading a class. If 90% of the class is good at math and 10% is struggling, a standard test might only reward the 90%. CW-B is like a teacher who says, "I care twice as much about the struggling students." If the computer misses a rare, high-risk patient, it gets a huge "penalty." If it misses a common patient, the penalty is smaller. This forces the computer to pay attention to the rare, dangerous cases.

3. The "Auditor" (Error Checking)

The program doesn't just guess; it keeps a ledger of its mistakes. It specifically checks: "Did we miss any of the high-priority groups (Stable, Emergency, and Suspected)?"

Analogy: It's like a security guard who doesn't just count how many people passed through the door, but specifically checks if they missed any VIPs or dangerous individuals.

How It Was Tested

The researchers tested CW-B against other smart programs (like standard AI, deep learning neural networks, and other sorting tools) using real hospital data from 4,354 patients.

The Results:

CW-B won. It was the best at getting the overall classification right.
It was the fairest. It had the best score at correctly identifying the rare groups (Macro-F1).
It was the safest. It made the fewest mistakes on the "high-priority" groups that doctors care about most.
It was transparent. Unlike some "black box" AI that you can't understand, CW-B uses "decision trees" (like a flowchart). You can trace exactly why it put a patient in a specific group, which is crucial for doctors to trust the system.

The Bottom Line

The paper claims that CW-B is a practical tool for hospitals. It takes messy, incomplete real-world data and sorts patients into the right care groups better than other methods, specifically ensuring that dangerous or rare conditions aren't overlooked. It does this by treating missing data as a clue, forcing the computer to care about rare cases, and keeping a clear record of its logic so doctors can audit it.

Technical Summary: CW-B for Imbalance Resilient Multi-Class Cardiac Phenotyping

Problem Definition
The paper addresses the challenge of automated cardiac discharge phenotyping, a task critical for bridging in-hospital care with post-discharge management, risk stratification, and follow-up planning. The authors identify two primary obstacles in real-world clinical data:

Class Imbalance: Real-world records exhibit skewed distributions where high-risk or clinically actionable phenotypes are often underrepresented compared to stable or non-disease states.
Data Incompleteness: Clinical records frequently contain missing values due to incomplete documentation or variable measurement protocols.

Standard machine learning approaches often optimize global metrics (e.g., overall accuracy) at the expense of clinically salient but rare phenotypes. Furthermore, the consequences of errors are asymmetric; missing a high-risk acute phenotype (e.g., Acute Coronary Syndrome) can lead to severe patient harm, whereas false positives may result in unnecessary but less critical follow-up. Existing methods, including rule-based systems, conventional ensembles, and deep learning models, struggle to balance generalizability, interpretability, and explicit handling of these asymmetric risks and missing data patterns.

Methodology: The CW-B Framework
The authors propose CW-B, a clinical risk-aligned class-weighted XGBoost pipeline designed for a five-class cardiac discharge phenotyping task. The framework integrates three core components to address imbalance and missingness without sacrificing interpretability:

Task Formulation:
The problem is defined as a multi-class classification with five labels: stableCAD (0), ACS (1), oldMI (2), CAS (3), and nonCAD (4).
- Clinically Prioritized Set ( $C_p$ ): The authors define a subset $C_p = \{0, 1, 3\}$ (stableCAD, ACS, CAS) based on clinical actionability and risk. These classes represent phenotypes with established diagnostic and therapeutic pathways. The framework explicitly audits errors within this set, distinct from global metrics.
Explicit Missingness Modeling:
Rather than treating missing data as a preprocessing nuisance, CW-B treats missingness as an informative feature.
- Leakage-Free Protocol: Preprocessing statistics (mean, standard deviation, median) are computed strictly on the training fold and applied to the validation fold to prevent data leakage.
- Augmentation: Missing entries are imputed using the training median. Crucially, a binary indicator vector ( $r$ ) is concatenated to the feature vector, encoding whether a value was originally observed or missing. This allows the model to distinguish between an imputed value and the clinical fact of unrecorded data.
Class-Balanced Instance Weighting:
To mitigate label imbalance, CW-B employs fold-specific instance weighting within the XGBoost objective function.
- Weight Calculation: For each class $c$ , a weight $w_c$ is calculated as the ratio of the total training samples to the number of samples in class $c$ . These weights are normalized to ensure stability.
- Integration: These weights are assigned to individual training instances ( $s_i = \tilde{w}_{y_i}$ ) and modulate the gradient boosting process, influencing both split selection and leaf value estimation. This reduces the dominance of frequent phenotypes (e.g., stableCAD) and forces the model to pay more attention to minority classes.

Key Contributions
The paper outlines three specific contributions:

Risk-Aligned Formulation: Framing cardiac discharge phenotyping as a task that explicitly emphasizes clinically actionable missed detections within a unified five-class objective.
Integrated Pipeline: Implementing a leakage-free, fold-specific class-balanced instance weighting scheme combined with explicit missingness-aware feature construction.
Comprehensive Benchmarking: Evaluating the proposed method against tree-based, ensemble, and neural baselines using both global metrics and prioritized class-wise error auditing.

Experimental Results
The framework was evaluated on a dataset of 4,354 samples with 57 structured features using stratified five-fold cross-validation.

Performance: CW-B achieved the highest Accuracy (0.81), Macro-F1 (0.72), Balanced Accuracy (0.73), and Prioritized F1 (0.69) among all compared methods.
Comparisons:
- CW-B outperformed a higher-capacity XGBoost baseline (more iterations, deeper trees), indicating that gains were due to the weighting strategy rather than model capacity.
- It surpassed CatBoost (CB), Stacking Ensembles (STK), and various Neural baselines (DQN, BC, MLP).
- While the Stacking Ensemble achieved the lowest prioritized miss rate, it did so at the cost of overall accuracy and macro-level balance. CW-B offered the best trade-off, delivering the highest prioritized F1 while maintaining strong global performance.
Ablation Studies: Removing class-balanced weighting or missingness indicators resulted in significant drops in Macro-F1, Balanced Accuracy, and Prioritized F1, confirming the necessity of both components.

Significance and Claims
The authors claim that CW-B provides a practical and deployment-oriented approach for reliable cardiac discharge phenotyping in real-world clinical settings. The significance of the work lies in:

Clinical Alignment: By explicitly modeling class bias and asymmetric risk, the framework reduces the likelihood of missing high-risk phenotypes that could divert subsequent management.
Interpretability and Auditability: Unlike deep learning models, the tree-based structure of CW-B preserves feature-based split structures. This supports traceable decision rationales and systematic auditing, which are essential for building clinical trust and facilitating adoption in regulated healthcare environments.
Resilience: The explicit modeling of missingness allows the system to handle the incomplete documentation and measurement variability inherent in real-world electronic health records (EHRs) without relying on complex imputation heuristics that might obscure the nature of the missing data.

The paper concludes that CW-B offers a reproducible, evidence-based framework that balances high global discrimination with the specific need to audit and reduce high-risk missed detections, directly supporting secondary prevention and consistent risk-aligned care.