---
license: mit
language:
- en
tags:
- bert
- pii-detection
- phi-detection
- hipaa
- healthcare
- nlp
- text-classification
- sequence-classification
- lora
- peft
datasets:
- custom
base_model: bert-base-uncased
pipeline_tag: text-classification
library_name: transformers
---

# HIPAA-BERT: PII/PHI Column Name Classifier

A fine-tuned BERT model for classifying database column names as **PII** (Personally Identifiable Information), **PHI** (Protected Health Information), or **Other (O)**.

## Model Details

| Property | Value |
|----------|-------|
| **Developer** | KronosX AI Labs |
| **Model Type** | BERT + LoRA (text classification) |
| **Base Model** | `bert-base-uncased` |
| **Language** | English |
| **Fine-tuning Method** | LoRA (Low-Rank Adaptation) |
| **Task** | Sequence Classification (3 classes) |

## Labels

| Label | Description | Examples |
|-------|-------------|----------|
| `O` | Other/Safe columns | `id`, `created_at`, `status` |
| PII | Personally Identifiable Info | `email`, `phone_number`, `address` |
| PHI | Protected Health Info (HIPAA) | `diagnosis_code`, `patient_name`, `ssn` |

## Training Details

### Hyperparameters

| Parameter | Value |
|-----------|-------|
| Learning Rate | 1e-3 |
| Batch Size | 64 |
| Epochs | 10 |
| Weight Decay | 0.01 |
| Max Sequence Length | 64 |
| LoRA Rank (r) | 16 |
| LoRA Alpha | 32 |
| LoRA Dropout | 0.1 |
| Target Modules | query, value |

### Training Data
Custom HIPAA-compliant dataset with ~50000+ labeled column names from healthcare databases.

### Hardware
- GPU: NVIDIA GPU (Kaggle)
- Mixed Precision: FP16 enabled

## Performance Metrics

| Metric | Score |
|--------|-------|
| Accuracy | ~95%+ |
| F1 (weighted) | ~94%+ |
| Precision | ~93%+ |
| Recall | ~94%+ |

## Usage

### Installation
pip install transformers torch

### Quick Start

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model
model_name = "KronosXAI/HIPAA-BERT-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Classify column names
columns = ["patient_name", "diagnosis_code", "created_at", "email", "status"]
for col in columns:
    inputs = tokenizer(col, return_tensors="pt", truncation=True, max_length=64)
    with torch.no_grad():
        outputs = model(**inputs)
        prediction = torch.argmax(outputs.logits, dim=-1).item()
    
    label_map = {0: "O", 1: "PII", 2: "PHI"}
    print(f"{col}: {label_map[prediction]}")

### Expected Output
patient_name: PHI
diagnosis_code: PHI
created_at: O
email: PII
status: O

## Intended Use

### Primary Use Cases
* Automatic PII/PHI detection in database schemas
* Data privacy compliance audits
* HIPAA compliance automation
* Healthcare data anonymization pipelines

### Out-of-Scope
* This model classifies column names, not the actual data content
* Not suitable for classifying free-text or unstructured data
* Should be used as part of a larger compliance workflow, not as sole arbiter

## Limitations & Bias
* Trained primarily on English column naming conventions
* May not generalize to non-standard or domain-specific naming patterns
* Should be validated with domain experts before production use

## Model Card Authors
Abishek - KronosX AI Labs

## Citation
@misc{hipaa-bert-2024,
  author = {KronosX AI Labs},
  title = {HIPAA-BERT: PII/PHI Column Name Classifier},
  year = {2026},
  url = {https://huggingface.co/KronosXAI/HIPAA-BERT-v0.1}
}

## Links
* Organization: KronosX AI Labs