medshield-py 0.1.0

Privacy-first PII/PHI sanitization for Medical AI Agents (Python Version).

MedShield-Py 🛡️

Professional PII/PHI Sanitization for Medical AI Agents (Python Version)

MedShield-Py is a high-performance, privacy-first library designed to protect sensitive patient data in AI-driven healthcare workflows. It mirrors the architecture of medshield-js, providing a consistent privacy layer across your entire stack.

🚀 Key Features

• Deterministic Tokenization: Replaces sensitive data (e.g., John Doe) with persistent IDs (e.g., [PERSON_0]) to maintain clinical context across conversation turns.
• Medical Context Preservation: Automatically detects and preserves medical codes (ICD-10, RXNORM) while scrubbing PII/PHI.
• Agentic Middleware: Drop-in integrations for LangChain, OpenAI, and newer AI frameworks.
• Synthetic Data Engine: Replaces real data with medically plausible fake data for developer environments.

📦 Installation

pip install medshield-py

🛠️ Quick Start

Basic Sanitization

from medshield import sanitize

text = "Patient John Doe (SSN: 123-45-6789) has asthma."
clean_text = sanitize(text)
# Output: "Patient [PERSON_0] (SSN: [SSN_0]) has asthma."

LangChain Integration

from medshield.integrations.langchain import MedShieldTransformer
from langchain_community.document_loaders import TextLoader

loader = TextLoader("patient_record.txt")
docs = loader.load()

# Sanitize all documents before indexing into a Vector DB
sanitizer = MedShieldTransformer(options={"level": "SYNTHESIZE"})
sanitized_docs = sanitizer.transform_documents(docs)

Managed Session (Deterministic)

from medshield import Sanitizer

s = Sanitizer(options={"level": "MASK"})
print(s.sanitize("Call John at 555-0101")) # [PERSON_0] at [PHONE_0]
print(s.sanitize("Is John home?"))        # Is [PERSON_0] home? (Identifies same person)

🛡️ Sanitization Levels

• MASK (Default): Replaces data with consistent tokens like [PERSON_0].
• SYNTHESIZE: Replaces real data with plausible fakes (e.g., "John Doe" -> "Robert Smith").
• REDACT: Simple redaction with type labels like [PERSON].

📜 License

Apache License 2.0