🚗 EDR Forensic Reconstruction & Crash Severity Prediction System

An industrial-grade forensic machine learning & deep learning system designed to reconstruct automotive crash severity (Deploy, Near-Deploy, Non-Deploy) using pre-impact telemetry logged across 46,533 NHTSA Event Data Recorder (EDR) records (2017–2024).

---

📌 Executive Summary & Key Highlights

• 🔒 Zero Data Leakage Protocol: Features a strict Train-First MICE Imputation pipeline (IterativeImputer with BayesianRidge), fitting on 80% Training set ONLY to transform 20% Test data cleanly.
• ⚖️ Imbalanced Class Learning (3 Pillars): Combines SMOTE oversampling ($k=3$), cost-sensitive tree class weighting (class_weight='balanced'), and PyTorch Focal Loss ($\gamma=2.0$).
• 🏆 Multi-Model Benchmark Leaderboard:
  • 🥇 XGBoost Classifier: 0.8832 Macro F1 | 97.28% Accuracy | 0.0502 MAE
  • 🥈 LightGBM Classifier: 0.8674 Macro F1 | 0.9932 AUC-ROC | SHAP Interpretability
  • 🤖 FT-Transformer (PyTorch): 65.82% Near-Deploy Recall (Highest minority sensitivity)
• 📦 ONNX Cross-Platform Runtime: All 6 trained models are exported to standard Open Neural Network Exchange (ONNX) format with 100.00% prediction class parity against native format binaries.
• 🛡️ Physical NHTSA Simulation Validation: Validated against 4 real-world NHTSA crash scenarios with 100% empirical ground truth agreement.

---

📊 Performance Benchmark Leaderboard

Model Architecture	Macro F1 (Primary)	Near-Deploy Recall	Accuracy	MAE (Lower is Better)	RMSE	$R^2$ Score	AUC-ROC
⚡ XGBoost Classifier	0.8832	60.76%	97.28%	0.0502	0.3101	0.8848	0.9918
🟢 LightGBM Classifier	0.8674	62.03%	97.14%	0.0519	0.3139	0.8820	0.9932
🌲 Random Forest	0.8646	58.23%	96.77%	0.0597	0.3384	0.8628	0.9869
🐱 CatBoost Classifier	0.8562	60.76%	97.18%	0.0503	0.3075	0.8867	0.9921
🤖 FT-Transformer (PyTorch)	0.8166	65.82% ⭐	96.60%	0.0576	0.3243	0.8740	0.9760
🪽 SAINT Transformer (PyTorch)	0.7942	62.03%	94.28%	0.1040	0.4446	0.7632	0.9757

---

🚀 Quick Start Guide

1. Installation

Clone the repository and install required dependencies:

git clone https://github.com/Byakuya3456/EDR_Project.git
cd EDR_Project
pip install -r requirements.txt

2. Execution & Notebook Inspection

Open and run the primary executed notebook:

jupyter lab EDR_Pipeline_Executed.ipynb

3. ONNX Runtime Inference Example

Run inference on saved ONNX models in Python, C++, or C#:

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("onnx_models/xgboost_model.onnx")
input_name = session.get_inputs()[0].name
predictions = session.run(None, {input_name: sample_features.astype(np.float32)})

---

📑 How to Cite This Work

If you reference, view, or cite this codebase, ONNX models, dataset pipeline, or forensic reconstruction methodology in academic research, papers, or projects, you MUST provide proper attribution:

@misc{srijith2026edr,
  author = {SRIJITH},
  title = {EDR Forensic Reconstruction & Crash Severity Prediction System},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/Byakuya3456/EDR_Project}},
  note = {Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)}
}

---

📁 Repository Structure

EDR_Project/
├── LICENSE                                # CC BY-NC-ND 4.0 Public License
├── README.md                              # Main Project Overview & Benchmark Documentation
├── requirements.txt                       # Python Dependencies Manifest
├── EDR_Pipeline.ipynb                     # Clean Primary Pipeline Notebook Scaffold
├── EDR_Pipeline_Executed.ipynb            # Fully Executed Notebook with Output Renderings
├── edr_pipeline_output/                   # Fused Master Dataset (46,533 records) & High-Res Graphics
├── onnx_models/                           # All 6 Exported ONNX Format Models & Parity Verification Report
├── project_report/                        # Detailed Research Reports & Metric Summaries
└── saved_models/                          # Native Serialized Model Weights & Preprocessors

---

📜 License & Copyright Protection

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).

To view a copy of this license, visit https://creativecommons.org/licenses/by-nc-nd/4.0/.

Summary of Legal Permissions:

• 👤 Attribution Required: You MUST give appropriate credit to SRIJITH, provide a link to the license, and indicate if changes were made.
• 🚫 NonCommercial: You MAY NOT use the material for commercial purposes or monetary gain.
• 🚫 NoDerivatives: If you remix, transform, or build upon the material, you MAY NOT distribute the modified material.

Copyright (c) 2026 SRIJITH. All Rights Reserved.