Artificial Intelligence for Liver Disease Prediction: A Systematic Review and an Explainable Hybrid Ensemble with Automated Clinical Feature Engineering

Technical Summary: Artificial Intelligence for Liver Disease Prediction

Problem Statement
The prediction of liver disease using routine clinical biomarkers, particularly on the Indian Liver Patient Dataset (ILPD), exhibits significant variability in reported performance. Literature reviews indicate headline accuracies ranging from approximately 71% to 99.6%. The paper posits that this dispersion is not solely due to predictive gains but stems from methodological artifacts, including improper data preprocessing, feature selection performed prior to validation, data leakage (e.g., duplicate rows crossing partitions, upsampling before splitting), and heterogeneous validation protocols. Furthermore, many studies prioritize raw accuracy over clinically relevant metrics like sensitivity, specificity, calibration, and robustness, leading to models that may appear effective but fail in practical screening scenarios.

Methodology
The study employs a dual approach: a DOI-verified systematic evidence review and a reproducible, leakage-aware experimental benchmark.

1. Systematic Review:

   • Scope: Synthesized structured data studies published between January 2020 and September 2026 with verified DOIs.
   • Focus: Explicitly analyzed validation protocols, class-balancing strategies, and resampling techniques to contextualize reported accuracy variations.
   • Findings: Identified that studies employing aggressive synthetic data workflows, pre-split upsampling, or non-nested feature selection often report inflated accuracies (up to 99.6%), whereas rigorous, leakage-aware protocols yield more modest results (71–74%).

2. Experimental Benchmark:

   • Dataset: The full 583-record ILPD (416 disease, 167 non-disease) was used. A sensitivity analysis was also conducted on a deduplicated 570-record cohort.
   • Data Quality & Preprocessing: Missing values were imputed using median values learned strictly from training folds. Duplicate rows were identified but retained for comparability, with a separate sensitivity analysis performed.
   • Automated Clinical Feature Engineering: A deterministic transformer expanded the original 10 predictors into a 24-feature candidate pool. This included clinically interpretable ratios (e.g., Direct/Total bilirubin, De Ritis ratio), composite biomarkers (e.g., bilirubin burden, enzyme sum), and log-transformations to address skewness.
   • Feature Selection: Mutual Information (MI) was used for feature ranking, strictly tuned within inner cross-validation folds to prevent data leakage.
   • Model Architecture: A hybrid ensemble was constructed using four complementary learners: Extra Trees, LightGBM, Gaussian Naive Bayes, and Logistic Regression. These were stacked using leakage-safe out-of-fold probabilities, followed by sigmoid calibration.
   • Validation Protocol: A nested 5×3-fold cross-validation was employed. The outer loop (5 folds) served as the test set, while the inner loop (3 folds) handled hyperparameter tuning, feature selection, and threshold optimization. No information from the outer test folds influenced preprocessing or model selection.
   • Threshold Policies: Two policies were evaluated: a "balanced-accuracy" threshold and a "screening" threshold (constrained to ≥80% sensitivity on training folds to maximize specificity).

Key Results

• Baseline Performance: Standard baseline models (e.g., SVM-RBF) demonstrated "majority-class collapse," achieving high sensitivity (1.0) but near-zero specificity (0.004), resulting in misleadingly high raw accuracy (~71.5%) while classifying almost all patients as diseased.
• Impact of Feature Engineering: The automated feature engineering module improved model discrimination.
  • ROC-AUC: Increased from 0.744 (original features) to 0.761 (engineered features).
  • PR-AUC: Improved from 0.893 to 0.897.
  • Balanced Accuracy: Remained stable at ~0.712 under the balanced threshold but shifted the trade-off toward better sensitivity.
• Threshold-Specific Performance:
  • Under a sensitivity-constrained threshold (aimed at screening), the engineered model achieved 69.5% accuracy, 78.1% sensitivity, 47.9% specificity, and an MCC of 0.258.
  • Under a balanced-accuracy threshold, the model achieved 64.8% accuracy, 56.2% sensitivity, and 86.2% specificity.
• Feature Stability: Nine features (including Age, ALT, AST, and engineered variables like indirect bilirubin and log-transformed aminotransferases) were selected in 100% of the nested-CV base-model fits, indicating robust signal detection across different model families.
• Explainability: SHAP analysis on the LightGBM base learner confirmed that Age, Direct Bilirubin, and protein ratios were primary drivers, aligning with clinical intuition.

Significance and Claims
The paper claims that its primary contribution is not achieving a new state-of-the-art accuracy, but rather demonstrating that protocol design fundamentally dictates the apparent difficulty and performance of the ILPD benchmark.

• Modesty in Claims: The author explicitly states that the proposed model does not exceed 90% accuracy, achieving instead a modest ROC-AUC of 0.761. They argue this is not a failure but a reflection of a rigorous, leakage-free validation design that prevents optimistic bias.
• Methodological Rigor: The study highlights that high headline accuracies in existing literature are often artifacts of methodological choices (e.g., resampling before splitting). The proposed "leakage-aware" framework provides a more credible, auditable, and defensible basis for medical AI research.
• Clinical Utility: The engineered ensemble improves the trade-off between sensitivity and specificity and offers better probability calibration (lower Brier score). However, the author cautions that the model is a "prototype research system" for decision support and screening, not a standalone clinical diagnostic tool, due to the dataset's small size, lack of external validation, and the high false-positive rate (specificity ~48%) at the screening threshold.
• Feature Engineering Value: The study validates that automated, clinically interpretable feature engineering (ratios, composites, log-transforms) can extract stable, predictive signals from small tabular medical datasets when integrated strictly within the training pipeline.

In conclusion, the paper asserts that robust feature engineering, strict leakage prevention, and transparent class-specific evaluation are essential for credible medical AI, and that future comparisons must account for validation design rather than relying on headline accuracy alone.