Explaining Individualized Treatment Rules: Integrating LIME and SHAP with XGBoost in Precision Medicine

Overview

XGBoostML extends the standard XGBoost framework by incorporating modified loss functions tailored for estimating individualized treatment rules (ITRs). This repository provides a suite of tools to improve the interpretability of ITR models using LIME and SHAP, along with statistical testing procedures to assess treatment effect heterogeneity.

Key Components

• XGBoostML_LIME
  Incorporates Local Interpretable Model-agnostic Explanations (LIME) into the XGBoostML framework to generate patient-specific explanations of treatment recommendations.

• XGBoostML_SHAP
  Integrates SHapley Additive exPlanations (SHAP) to attribute feature-level contributions to treatment assignments, both locally and globally.

• Global Permutation Test
  Implements a permutation-based hypothesis test to detect global treatment effect heterogeneity under the XGBoostML framework. This helps determine whether interpretability tools should be applied in a given dataset.

• Doubly Robust (DR) Modified Loss Functions
  Provides DR-enhanced versions of the modified loss function for continuous and binary outcomes in ITR estimation.

Getting Started

Main Notebooks for SHAP and LIME Interpretability

The following notebooks provide end-to-end implementations for different outcome types:

• XGBoostML_LIME_SHAP_continuous.ipynb – Continuous outcomes
• XGBoostML_LIME_SHAP_binary.ipynb – Binary outcomes
• XGBoostML_LIME_SHAP_time_to_event.ipynb – Time-to-event (survival) outcomes

Note: Each notebook uses a fixed set of hyperparameters for XGBoostML (not necessarily optimal). You are encouraged to tune hyperparameters based on your data and application.

Prerequisites

Please ensure the following Python packages are installed:

• Python 3.7+
• xgboost
• scikit-learn
• shap
• lime
• numpy
• pandas
• matplotlib