Explainable Ensemble-Based Machine Learning Models for Detecting the Presence of Cirrhosis in Hepatitis C Patients

Technical Summary: Explainable Ensemble-Based Machine Learning Models for Detecting Cirrhosis in Hepatitis C Patients

Problem Statement
Hepatitis C virus (HCV) infection is a global public health crisis, particularly in Egypt, where prevalence rates are among the highest worldwide. A critical complication of chronic HCV is liver cirrhosis, a condition characterized by permanent scarring that often progresses silently until advanced stages, leading to liver failure, gastrointestinal hemorrhage, and increased mortality. While direct-acting antivirals (DAAs) can cure HCV, many patients already suffer from chronic infection and associated complications. Current diagnostic methods for cirrhosis, such as liver biopsies, are invasive, expensive, and risky. Non-invasive alternatives like transient elastography or clinical indices (e.g., APRI, FIB-4) suffer from limitations in availability, accuracy, and applicability across diverse patient populations (e.g., those with ascites or obesity). Although machine learning (ML) has shown promise in diagnosing various diseases, there is a notable gap in research applying ML specifically to detect the presence of cirrhosis in HCV patients. Furthermore, the "black box" nature of complex ML models often hinders their adoption by medical professionals who require explainable decision-making processes.

Methodology
This study utilized a dataset of 2,038 Egyptian patients (derived from the UCI Machine Learning Repository) containing 28 attributes. The target variable was binarized to distinguish between patients with cirrhosis (positive class) and those with earlier stages of fibrosis or no cirrhosis (negative class).

The research followed a structured pipeline:

1. Preprocessing: The data underwent outlier removal using the Interquartile Range (IQR) method and normalization via Min-Max scaling. To address the class imbalance inherent in the dataset (where cirrhosis cases were fewer), random oversampling was applied.
2. Model Training: Four ensemble-based machine learning algorithms were trained and compared:
   • Random Forest (RF)
   • Gradient Boosting Machine (GBM)
   • Extreme Gradient Boosting (XGBoost)
   • Extra Trees (ET)
3. Optimization: Hyperparameters for each algorithm were tuned using GridSearchCV with stratified 10-fold cross-validation.
4. Feature Selection: Sequential Feature Selection (SFS) was employed to identify the most significant subset of features, aiming to reduce dimensionality and improve model efficiency.
5. Explainability: The best-performing model was analyzed using Explainable Artificial Intelligence (XAI) techniques, specifically Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), to interpret feature contributions and build trust in the model's decisions.

Key Contributions
The study makes four primary contributions to the field of medical data science:

• Comparative Analysis: It evaluates the performance of four distinct ensemble learning algorithms specifically for diagnosing cirrhosis in HCV patients, a specific application not previously explored in depth.
• Feature Reduction: It applies Sequential Feature Selection (SFS) to minimize the number of attributes required for accurate diagnosis, demonstrating that high performance can be achieved with a reduced feature set.
• Explainability Integration: It utilizes XAI techniques (SHAP and LIME) to demystify the model's decision-making process, addressing the trust barrier often associated with AI in healthcare.
• Identification of Key Biomarkers: It identifies the most significant clinical attributes for diagnosing cirrhosis in this specific cohort, offering insights into the physiological markers most predictive of the disease state.

Results
The study found that data balancing via random oversampling was critical; without it, boosting models (GBM and XGBoost) failed to predict positive cases (0% recall). After oversampling and hyperparameter tuning, all models showed significant improvement.

Following Sequential Feature Selection (SFS):

• Extra Trees (ET) emerged as the superior model. It achieved the highest accuracy of 96.92%, a precision of 99.81%, a recall of 94.00%, and an AUC-ROC of 96%. Notably, the ET model achieved this performance using only 16 out of the original 28 features.
• Random Forest (RF) followed closely with 96.58% accuracy.
• XGBoost achieved the highest recall (94.81%) but lower overall accuracy (90.99%) and precision (88.00%).
• GBM achieved 95.70% accuracy.

The confusion matrices indicated that while XGBoost had the fewest False Negatives (FN), the ET model had the fewest False Positives (FP), which is crucial for avoiding unnecessary medical interventions and patient anxiety.

Explainability Findings

• SHAP Analysis: Identified RNA 4, BMI, RNA 12, and AST 1 as the most important features contributing to predictions. High values in BMI and AST 1 showed positive contributions to the cirrhosis prediction, while high RNA 4 values showed negative contributions.
• LIME Analysis: For individual positive predictions, features such as BMI, RNA Base, ALT 48, RNA 4, and ALT 36 were highlighted as key drivers.

Significance and Claims
The authors claim that this study demonstrates the viability of using explainable ensemble learning to assist medical specialists in the early detection of cirrhosis in HCV patients. By achieving high accuracy with a reduced set of features, the proposed model could potentially reduce the need for extensive clinical testing, thereby optimizing time utilization for both professionals and patients and contributing to cost savings. The integration of XAI ensures that the model's outputs are interpretable, fostering trust among healthcare providers.

However, the paper maintains a modest stance regarding its broader implications. The authors acknowledge that the model was trained on synthetic data generated via oversampling rather than a larger, real-world dataset of cirrhosis patients. Consequently, they suggest that future research should focus on validating the model with real data and expanding the scope to predict other stages of HCV progression to create a more generalized diagnostic tool.