Federated Tree-Based Ensembles with SHAP Explainability and Integrated Feature Selection for Secure Lung Cancer Health Analytics

Aim/Purpose: Federated learning (FL) enables collaborative model training without centralizing raw data, thereby preserving privacy compared to traditional centralized machine learning approaches. To develop a privacy-preserving and explainable framework – where explainability is achieved through local SHAP analyses at each client and an aggregated global SHAP view – for health risk prediction in decentralized, heterogeneous environments, addressing the challenge of maintaining data confidentiality while achieving high predictive performance. Background: Traditional centralized health analytics solutions compromise data privacy; this paper introduces a federated learning approach that enables collaborative model building without exposing raw data. Methodology: The study integrates TreeNetX, a stable ensemble meta-learner, with SHAP-based explainability, leveraging client-specific training using gradient boosting models (XGBoost, LightGBM, CatBoost) and federated aggregation to address non-IID client distributions. Each client performs local feature selection using Recursive Feature Elimination with Cross-Validation (RFECV) before training. Experiments were conducted on three heterogeneous, tabular lung-cancer datasets from Kaggle to validate the model’s performance, using metrics including accuracy, precision, recall, F1-score, AUC, and calibration-related measures. Contribution: This paper contributes a novel federated ensemble learning framework that enhances predictive accuracy, robustness, and model interpretability, while ensuring data privacy in sensitive healthcare applications. Findings: The federated TreeNetX model achieved 91.17% accuracy and an AUC of 0.9667. It outperformed individual client models in robustness and generalization. SHAP-based analysis provided clinically meaningful insights, enhancing model trustworthiness. Recommendations for Practitioners: Practitioners should adopt federated ensemble strategies like TreeNetX for privacy-conscious predictive analytics to maintain compliance with data protection regulations while achieving high model performance. Recommendation for Researchers: Researchers should further explore hybrid federated architectures that combine explainability and advanced ensemble techniques to optimize interpretability and performance across heterogeneous environments. Impact on Society: The approach promotes ethical, privacy-preserving AI adoption in healthcare and other sensitive fields, contributing to safer and more trustworthy smart systems. Future Research: Future studies should investigate scaling the framework to even larger, more diverse client networks, integrating additional explainability tools, and extending applications beyond healthcare to other regulated industries.