Meddies Sprouts Meddies PII To Pluck Patient Data From Clinical Text Across 17 Languages

    
        By vramkickedin    
     | 
    
            June 22, 2026 at 10:15 pm        
    
     | 
    
        2 min read

Meddies PII is a new open multilingual model that extracts personally identifiable information (PII) from clinical and administrative text, returning structured JSON instead of raw token tags. It identifies seven types of identifiers across 17 languages, such as names, phone numbers, email addresses, physical addresses, dates, ID numbers, and company names. The release is a research artifact meant for privacy and healthcare AI teams, not a redaction tool or a finished compliance product.

Meddies, a small team focused on verifiable clinical intelligence, built this extractor to tackle the problem of safely pulling PII without erasing the clinical facts that downstream systems still need. Traditional NER pipelines often break on token alignment, language-specific span quirks, and brittle post-processing when document formats shift, so the group opted for a causal language model that accepts chat-style input and outputs deterministic JSON. By pairing a synthetic dataset with dynamic prompting, they aimed for stable extraction behavior across many languages, document structures, and runtime environments.

Multilingual extraction with deterministic JSON output

Key Features

Extracts 7 PII entity types into JSON.
Handles 17 languages with consistent behavior.
Runs on consumer GPUs and in browsers.
Accepts raw text via chat-style prompts.
Deterministic generation with temperature zero.
Set-based exact-match scoring for accuracy.

Healthcare developers and researchers can use the model to de-identify clinical notes, discharge summaries, and admin forms in multiple languages without destroying useful medical details. Small teams running local AI on consumer hardware will find it practical thanks to a modest GPU footprint and ONNX variants that deploy directly in browsers or edge devices. Privacy-focused professionals can treat it as a structured extraction baseline, adding their own policy, redaction, and compliance validation afterward.

Performance notes and model limitations

The model reaches an overall entity F1 of 0.81 on the evaluation set, with phone numbers and email addresses scoring above 0.92, while the company name class lags at 0.33 due to a label-definition mismatch. Around 1.3% of the generated values are hallucinations that do not appear in the input, and the system only extracts identifiers — it does not redact or anonymize text on its own.

Future improvements may address weaker entity classes and nested reasoning, but the team cautions that solid benchmark numbers do not guarantee GDPR, HIPAA, or other regulatory compliance on real data.