CrashPulse-AI

Early occupant chest-response prediction from public NHTSA crash-test data

This project is a public-data passive-safety machine-learning demonstrator.
It uses NHTSA frontal crash-test data to extract driver chest displacement and vehicle crash-pulse features, then builds models for occupant chest-response prediction and early high-response flagging.

The project was developed as a portfolio/research demonstrator for vehicle safety, crash-pulse analysis, occupant response prediction, and AI-based decision support.

---

Project goal

The main question is:

Can early vehicle crash-pulse information help identify frontal crash tests with elevated driver chest response?

The final early classifier uses only vehicle metadata and the first 0–30 ms of crash-pulse information.

---

Important disclaimer

This project does not claim to predict real medical injury risk.

The target used here is:

peak_abs_driver_chest_displacement_mm

This is a crash-test dummy response metric, not a direct injury label.

The project is:

• not a validated injury-prediction model,
• not a validated restraint-control algorithm,
• not based on proprietary OEM data,
• an open-data demonstrator for signal extraction, feature engineering, ML modeling, and conceptual decision support.

---

Workflow

NHTSA frontal crash-test data
→ TDMS signal extraction
→ driver chest displacement target
→ vehicle crash-pulse feature engineering
→ regression modeling
→ high chest-response classification
→ early 0–30 ms conceptual decision-support layer

---

Final results

Task	Final model	Feature set	Main result
Chest displacement regression	Random Forest	Advanced vehicle metadata + 0–30 ms pulse features	MAE ≈ 3.72 mm
Early high-response classification	Tuned Logistic Regression	Vehicle metadata + simple 0–30 ms pulse features	Recall ≈ 0.71, precision ≈ 0.22
Decision support	Tuned early 0–30 ms decision layer	Threshold = 0.45	High-response flag group ≈ 6.1× higher high-response rate than standard monitoring

High chest response was defined as:

peak driver chest displacement >= 30 mm

---

Key finding

The final tuned early model detected 35 out of 49 high chest-response cases using only the first 0–30 ms of crash-pulse information.

The final decision-level rates were:

Decision level	Tests	Actual high cases	High-response rate
Standard monitoring	255	9	3.53%
Elevated monitoring	60	5	8.33%
Early high-response flag	162	35	21.60%

This means the early high-response flag group had approximately 6.1× higher high-response rate than the standard-monitoring group.

---

Repository structure

Suggested structure:

CrashPulse-AI/
├── data/
│   └── processed/
├── notebooks/
│   └── CrashPulse_AI_GitHub_Clean_Notebook.ipynb
├── results/
│   ├── tables/
│   └── figures/
├── README.md
├── requirements.txt
└── .gitignore

Large raw TDMS/zip files should not be uploaded to GitHub.

---

Main notebook

The cleaned notebook is:

notebooks/CrashPulse_AI_GitHub_Clean_Notebook.ipynb

It focuses on the final cleaned project story:

1. project idea,
2. dataset overview,
3. chest displacement signal-quality check,
4. regression result,
5. high-response classification,
6. final tuned early decision-support layer,
7. limitations and future work.

---

Future work

Possible next steps:

1. Scale from 500 attempted tests to a larger NHTSA frontal-test dataset.
2. Use grouped validation by vehicle model/family.
3. Extract richer restraint-system, airbag, belt, dummy, seat, and occupant-position metadata.
4. Test raw-pulse time-series models once the dataset is larger.
5. Explore deep learning models such as 1D CNNs, TCNs, or tabular foundation models when enough data is available.
6. Apply the same workflow to HBM simulation data if such data becomes available.

---

Skills demonstrated

• Crash-test data handling
• NHTSA API/TDMS workflow
• Signal extraction and filtering sensitivity checks
• Crash-pulse feature engineering
• Regression modeling
• Imbalanced classification
• Threshold tuning
• Early decision-support logic
• Passive-safety interpretation