ZooStack: SHAP-Guided Meta-Ensemble Learning for Explainable Few-Shot Disease Prediction with Minimal Labeled Data

The adoption of machine-learning models in clinical decision support is often limited by scarce labeled data and weak few-shot transfer across heterogeneous datasets under controlled evaluation settings. This paper presents ZooStack, an explainable meta-learning ensemble framework for few-shot disease prediction that rapidly adapts across clinical datasets. ZooStack integrates heterogeneous base learners—Random Forest, Logistic Regression, XGBoost, and SVM—trained on structured electronic health record (EHR) features together with MedBERT-derived text embeddings, and combines them via a task-adaptive meta-learner. To improve stability and interpretability under data scarcity, we introduce SHAP-informed priors that regularize the meta-learner toward clinically meaningful features, mitigating overfitting in low-sample regimes. We meta-train on three disease tasks (Diabetes, Heart Disease, and Parkinson's) and evaluate generalization by cross-validating on a held-out task (Breast Cancer). Experiments show consistent gains from meta-learning and further improvements with SHAP regularization. On the Breast Cancer benchmark, ZooStack achieves 98.2% accuracy and 0.99 AUC after SHAP-guided meta-optimization. Few-shot adaptation improves substantially after optimization: 0-shot accuracy increases from 38.1% to 78.9%, 5-shot from 55.1% to 91.6%, and 10-shot from 59.4% to 98.2%. Compared with state-of-the-art approaches that require full dataset training, ZooStack attains competitive performance using only few-shot support examples. SHAP-based analyses further confirm preserved interpretability during adaptation, with clinically relevant features (e.g., glucose and BMI for diabetes; morphological descriptors for cancer) consistently receiving high attribution. These results suggest that explainability-driven meta-ensembles can enable practical few-shot clinical decision support without sacrificing transparency.