Data-driven assessment of cross-scale neurobiological risks from electromagnetic fields using interpretable machine learning

This repository contains the comprehensive cross-scale dataset and the unified machine learning source code supporting the manuscript: "Data-driven assessment of cross-scale neurobiological risks from electromagnetic fields using interpretable machine learning", submitted to the Journal of the Royal Society Interface. 1. Dataset Overview The dataset comprises 3,056 independent experimental and epidemiological samples meticulously extracted and structured from peer-reviewed literature. It spans three distinct biological scales: Cellular level: 1,548 samples Animal level: 1,015 samples Human level: 493 samples The structured data matrices encompass physically informed exposure features (including electric field, magnetic field, specific absorption rate, and time-integrated cumulative dose) and binary biological effect labels (significant/non-significant). All physical parameters have been standardized and processed to resolve dimensionality and multicollinearity issues. 2. Source Code (paper main.py) The provided Python script constitutes the complete, reproducible machine learning pipeline utilized in the study. Key functionalities include: Algorithm Benchmarking: Automated 10-fold stratified cross-validation across multiple algorithms (Logistic Regression, SVM, MLP, Random Forest, XGBoost, LightGBM, and Stacking Ensembles) using a custom Composite Score (0.60 × AUC + 0.40 × Sensitivity). OOF Threshold Optimization: Implementation of an Out-Of-Fold dynamic threshold optimization strategy to maximize screening sensitivity under a strict specificity baseline. SHAP Mechanism Decoding: Integration with the SHAP (SHapley Additive exPlanations) framework to calculate global feature importance and directional impacts. Phase Transition Extraction: A novel zero-crossing algorithm to decode complex nonlinear dynamics into explicit, quantitative phase transition thresholds (DRIP) for cross-scale safety evaluation. Automated Visualization: Code to automatically reproduce all high-resolution figures (ROC/PR curves, SHAP beeswarm plots, and dependence plots) and quantitative tables presented in the manuscript. 3. Usage Instructions To reproduce the analytical framework, ensure the necessary Python dependencies (e.g., scikit-learn, xgboost, lightgbm, shap, pandas, seaborn) are installed. Place the dataset files and the Python script in the designated working directory as configured in the script, and execute Environmental Research.py.