LLM BERT Model for HIPAA-Sensitive Database Fields Classification
This repository hosts a fine-tuned BERT-base model that classifies database column names as either PHI HIPAA-sensitive (e.g., birthDate, ssn, address) or non-sensitive (e.g., color, food, country).
Use this model for:
Masking PHI data fields before sharing database to avoid HIPAA compliance
Preprocessing before data anonymization
Identifying patient's sensitive data fields in a dataset before training an AI model
Enhancing security in healthcare and mHealth applications
🧠 Model Info
Base Model: bert-base-uncased
Task: Binary classification (PHI HIPAA Sensitive vs Non-sensitive)
Trained On: GAN generated Synthetic and real-world column name examples
In the healthcare industry, safeguarding sensitive patient data is of utmost importance, particularly when developing and maintaining software systems that involve database sharing. The Health Insurance Portability and Accountability Act (HIPAA) mandates strict regulations to ensure the privacy and security of Protected Health Information (PHI). Healthcare organizations must comply with these regulations to prevent unauthorized access, breaches, and potential legal consequences. However, ensuring HIPAA compliance becomes a complex challenge when databases are shared among multiple teams for debugging, development, and testing purposes.
This research work proposes a novel approach that uses BERT based LLM for identifying sensitive database columns into the database schema in order to avoid PHI HIPAA violation.
Disclaimer
This LLM model is fine-tuned with synthetic dataset(~5K) and is provided for research and educational purposes only. Always verify compliance before using in production environments.
📊 Model Performance Analysis
Table 1: Changing hyperparameters and results
Step
Learning Rate
Batch Size
Epoch
Weight Decay
Precision
Recall
F1 Score
Accuracy
1
0
16
1
0.001
0.0000
0.0000
0.0000
36.78%
2
1e-1
16
1
0.001
0.6321
1.0000
0.7746
63.21%
3
1e-1
32
1
0.001
0.6321
1.0000
0.7746
63.21%
4
1e-1
32
2
0.001
0.6321
1.0000
0.7746
63.21%
5
1e-1
32
3
0.001
0.6321
1.0000
0.7746
63.21%
6
1e-1
32
3
0.01
0.6321
1.0000
0.7746
63.21%
7
2e-1
32
4
0.01
0.6321
1.0000
0.7746
63.21%
8
3e-4
32
4
0.01
0.6331
0.9982
0.7748
63.32%
9
2e-4
32
4
0.01
0.9908
0.9730
0.9818
97.72%
10
1e-5
32
4
0.01
0.9964
0.9928
0.9946
99.31%
11
1e-5
32
5
0.01
0.9964
0.9928
0.9946
99.31%
12
1e-5
16
5
0.01
1.0000
0.9964
0.9982
99.72%
13
1e-5
16
5
0.1
1.0000
0.9946
0.9973
99.65%
14
1e-5
32
5
0.1
1.0000
0.9946
0.9973
99.65%
15
1e-5
32
5
1.0
0.9964
0.9946
0.9946
99.54%
16
1e-6
32
5
1.0
0.8342
0.9153
0.8729
83.15%
Limitations
One of the main limitations of this work is the use of
a synthetic dataset instead of real-world data to fine-tune
and train the AI models. Although the dataset was carefully
checked for accuracy, it may not fully reflect the complexity
and diversity of actual healthcare records.
👤 Author
MD Abdul Barek
PhD student & GRA @ Intelligent Systems and Robotics
Advisor:
Dr. Hakki Erhan Sevil
Associate Professor
Intelligent Systems and Robotics,
University of West Florida
📧 hsevil@uwf.edu
Supervisors:
Dr. Guillermo Francia III
Director, Research and Innovation,
Center for Cybersecurity,
University of West Florida
📧 gfranciaiii@uwf.edu
Dr. Hossain Shahriar
Associate Director and Professor, Center for Cybersecurity,
University of West Florida
📧 hshahriar@uwf.edu
Dr. Sheikh Iqbal Ahamed
Wehr Professor and Founding Chair of Computer Science Department at Marquette University,
Marquette University
📧 sheikh.ahamed@marquette.edu
