Local PHI Scrubber Chrome Extension

Chrome extension that removes Protected Health Information (PHI) from clinical text using a locally hosted Microsoft Phi-3 Mini 128K large language model served via Ollama. All data stays on your machine.

IMPORTANT CAVEAT: Language models are non-deterministic and may miss identifiers or over-redact. Manual review of all output is mandatory before use.

Demo

demo.mp4

How It Works

User Workflow

┌─────────────────────────────────────────────────────────────┐
│ 1. User enters clinical text in extension popup             │
└────────────────┬────────────────────────────────────────────┘
                 │
                 ▼
┌─────────────────────────────────────────────────────────────┐
│ 2. Extension sends text + prompt to local Ollama server     │
│    (HTTP POST to localhost:11434/api/generate)              │
└────────────────┬────────────────────────────────────────────┘
                 │
                 ▼
┌─────────────────────────────────────────────────────────────┐
│ 3. Phi-3 Mini processes text locally on user's machine      │
│    - Identifies PHI using Safe Harbor rules                 │
│    - Replaces with [NAME], [DATE], [MRN], etc.              │
│    - Returns de-identified text                             │
└────────────────┬────────────────────────────────────────────┘
                 │
                 ▼
┌─────────────────────────────────────────────────────────────┐
│ 4. Extension displays results for human review              │
│    - Shows original and scrubbed text comparison            │
│    - Counts PHI identifiers removed                         │
│    - **Mandatory manual verification step**                 │
└────────────────┬────────────────────────────────────────────┘
                 │
        ┌────────┴────────┐
        ▼                 ▼
┌──────────────────┐  ┌────────────────────────────────┐
│ Copy to Clipboard│  │ Inject into Any Online LLM     │
│ for manual paste │  │ (ChatGPT, Claude, Gemini, etc) │
└──────────────────┘  └────────────────────────────────┘

Project Structure

local-phi-scrubber/
├── manifest.json          # Chrome extension manifest (defines permissions, content scripts, icons)
├── popup.html             # Extension popup UI structure (HTML markup for the interface)
├── popup.js               # Main popup logic (handles user input, Ollama API calls, UI updates)
│                          # CONTAINS SYSTEM PROMPT: PROMPT_TEMPLATE constant (lines 11-82)
├── content.js             # Content script (injects scrubbed text into chat interfaces)
├── styles.css             # Popup UI styling (CSS for extension interface)
├── README.md              # Project documentation (this file)
├── LICENSE                # MIT license text
├── demo.mp4               # Demo video showing extension in action
└── icons/                 # Extension icon assets
    ├── icon16.png         # 16x16 icon (toolbar)
    ├── icon48.png         # 48x48 icon (extension management)
    └── icon128.png        # 128x128 icon (Chrome Web Store)

Installation & Setup

Prerequisites

Before installing the extension, ensure you have:

• Chrome or Edge version 128 or later
• Ollama installed on your machine
• 4GB+ system RAM minimum (8GB+ recommended for comfortable use)
• Optional but beneficial: GPU with VRAM for faster inference

Step 1: Install Ollama

Ollama is the local inference engine that runs Phi-3 Mini.

1. Visit https://ollama.ai and download Ollama for your operating system
   • macOS (Intel & Apple Silicon)
   • Windows
   • Linux
2. Install and launch Ollama
3. Ollama will run as a background service on http://localhost:11434

Step 2: Download Phi-3 Mini Model

Once Ollama is running, open a terminal and pull the Phi-3 Mini model:

ollama pull phi3:mini

This downloads the model (~2.03GB with Q4_0 quantization). First download may take 3-10 minutes depending on internet speed. You can also use ollama pull phi3:medium (14B) or ollama pull phi3 (4K context) variants if you prefer.

Verify installation with:

ollama list

You should see phi3:mini in the list of available models.

Step 3: Load the Extension in Chrome

1. Open Chrome and navigate to chrome://extensions
2. Enable Developer mode (toggle in top right)
3. Click Load unpacked
4. Navigate to the local-phi-scrubber folder and select it
5. The extension should appear in your extensions list and toolbar

Step 4: Verify Ollama Connection

1. Click the extension icon in your Chrome toolbar
2. You should see a green status indicator: "AI ready • Ollama + Phi-3"
3. If you see an error, verify:
   • Ollama is running (ollama serve in terminal)
   • Phi-3 Mini is installed (ollama list)
   • Port 11434 is accessible

Step 5: Test De-Identification

1. Paste sample clinical text in the extension textarea:

   Patient John Smith, DOB 12/15/1985, MRN #123456
   Admitted to Boston General Hospital on 3/10/2024
   Chief complaint: chest pain
   Contact: (617) 555-1234
   

2. Click "Scrub PHI"
3. Review the output (should show de-identified text with [NAME], [DATE], [MRN], etc.)
4. IMPORTANT: Manually verify the results before use in any external system

---

Technical Flow

1. User Input → Textarea in popup.html
2. Scrub Button Click → popup.js:scrubPHI() function
3. API Call → HTTP POST to localhost:11434/api/generate with:
   • Model: phi3:mini
   • Prompt: Complete HIPAA de-identification instructions
   • Query: User's clinical text
   • Temperature: 0 (for consistency)
4. Local Processing → Ollama runs Phi-3 Mini on user's hardware
5. Response → De-identified text returned
6. Analysis → Count and display identified PHI tokens
7. User Review → Manual verification required
8. Output → Either copy to clipboard or inject into chat interface

---

Features

What This Extension Provides

Feature	Description
AI-Assisted De-Identification	Uses Phi-3 Mini LLM for intelligent PHI removal
Local Processing	All computation happens on your machine, zero cloud transmission
HIPAA Safe Harbor Aligned	Targets all 18 Safe Harbor identifiers
Two Output Methods	Copy to clipboard OR inject directly into any online LLM platform
Real-Time Status	Shows Ollama connection status in UI
PHI Analytics	Displays count of identifiers removed
Minimal Permissions	Only requests necessary Chrome permissions
Open Source	MIT licensed, fully auditable code

What This Extension DOES NOT Provide

Limitation	Why It Matters
Guarantee of de-identification	LLMs are non-deterministic; manual verification is required
Standalone HIPAA compliance	Human review is mandatory; this is a tool, not a safeguard
100% PHI detection	Language models may miss some identifiers
Zero false positives	May over-identify and remove non-sensitive content
Real-time learning	Cannot improve from past errors without model retraining
Offline operation	Requires Ollama to be installed and running locally

---

Privacy & Security

What Data Stays Private

• All processing is local - Your text never leaves your machine
• No cloud API calls - No data transmission to external servers
• No telemetry - Extension doesn't track or report your usage
• No API keys needed - Ollama runs entirely on localhost
• No authentication required - No login, no accounts, no tracking
• Open source code - You can audit every line (1.2KB popup.js, 2.4KB content.js)
• MIT licensed - Permissive open source license, commercial use allowed

Permissions Explanation

The extension requests minimal permissions:

{
  "permissions": ["activeTab", "scripting"],
  "host_permissions": [
    "http://localhost:11434/*",        // Local Ollama inference
    "https://chatgpt.com/*",           // ChatGPT (OpenAI)
    "https://www.perplexity.ai/*",     // Perplexity AI
    "https://claude.ai/*",             // Claude (Anthropic)
    "https://gemini.google.com/*",     // Gemini (Google)
    "https://copilot.microsoft.com/*", // Copilot (Microsoft)
    "https://huggingface.co/*",        // Hugging Face Chat
    "https://github.com/*"             // GitHub Copilot
  ]
}

Note: You can add more LLM platforms by updating the host_permissions in manifest.json.

No tracking, no analytics, no third-party APIs.

---

System Requirements

Minimum Requirements (It's Smaller Than You Think!)

Component	Requirement
Browser	Chrome 128+ or Edge 128+
Operating System	Windows, macOS, Linux (any recent version)
System RAM	4GB minimum (8GB+ recommended)
Disk Space	2.5GB for Phi-3 Mini 128K model (Q4 quantized)
Internet	Only needed to download Ollama and model once

What Can Actually Run This

Hardware	Works?	Performance	Notes
5+ year old laptop (CPU)	Yes	5-10 tok/s	Q4 quantization, practical use
Modern laptop (CPU)	Yes	10-15 tok/s	No GPU needed
Laptop with GPU	Yes	20-40 tok/s	Much faster, optional
Desktop PC	Yes	Excellent	Best experience
Apple Silicon Mac	Yes	107 tok/s	Fastest consumer option
iPhone A16+	Yes	12 tok/s	ONNX optimized, practical
Android S21+	Yes	8+ tok/s	ONNX optimized, experimental
Raspberry Pi 4	Maybe	Very slow	Possible with extreme quantization

License

MIT License - This software is open source and free for commercial use.

See LICENSE file for complete license text.

---

Learning Resources

Understanding PHI & HIPAA Safe Harbor

• HIPAA De-Identification (Expert Determination & Safe Harbor) – HHS
• 18 HIPAA Safe Harbor Identifiers – 45 CFR § 164.514(b)(2)(i) (eCFR)

Understanding Small Language Models

• Phi-3 Technical Report – arXiv
• Phi-3 Mini 128k Instruct – Hugging Face
• Ollama Documentation

LLM Non-Determinism

• Softmax Function – Wikipedia
• Temperature and Sampling – OpenAI Docs