Medical Data Processing

Problems this solves

• Valuable information is buried in free-text clinical notes, case reports or publications.
• Records can't be used for research or AI until patient identifiers are removed.
• You need annotated medical text to train or evaluate a model.
• Drug names, conditions and adverse events are recorded inconsistently across sources.

For research and development. Our pipelines and datasets support research, evaluation and AI development. They are not medical devices and are not intended for diagnosis or treatment decisions. Any clinical use requires separate validation and, where applicable, regulatory approval.

What we do with medical data

Research data processing. Much of the most useful medical information lives in free text: case reports, clinical notes, adverse-event reports and published studies. We build pipelines that extract drugs, conditions, adverse reactions, dosages and patient characteristics into structured data, and can map terms to standard vocabularies such as ICD-10, SNOMED CT or MedDRA where your licenses permit.

De-identification. Before records can be used for research or AI development, identifying information usually has to be removed. We de-identify free text and structured records — names, dates, locations, record numbers and other identifiers — and measure how well it worked, rather than assuming it did.

Medical AI datasets. We design and annotate medical text datasets for training and evaluating models, with guidelines written for the clinical domain and labels reviewed by clinical experts from your team or advisors. See AI Datasets & Annotation for how we approach dataset quality.

De-identification done carefully

Automated de-identification is never perfect, so we treat it as a measured process:

• Measured accuracy. We test on a reviewed sample and report how many identifiers were caught and missed, so you can judge residual risk.
• Human review where it matters. Automated detection combined with human checking for high-risk material.
• The right standard for your context. In the United States, HIPAA defines de-identification through the Safe Harbor method (removing 18 categories of identifiers) or Expert Determination. Under the GDPR, health data is a special category, and pseudonymized data — where identities could be restored with additional information — is still personal data. We help you choose an approach that matches the standard you need to meet.

How we handle sensitive data

• Terms before data. Purpose, legal basis, access and retention are agreed in writing before any data is shared with us. Ethics approval and the legal basis for processing remain your responsibility as the data controller.
• Minimum necessary. We work with the smallest dataset that answers the question, de-identified at source whenever possible.
• Your environment if needed. Pipelines can run inside your infrastructure, using self-hosted language models, so data never leaves it. We never send medical data to external AI services without your explicit agreement.
• Documented and deleted. Every processing step is documented, and working copies are deleted at the end of the project as agreed.

Our medical data work is led by co-founder Shogher Ghazaryan, who spent six years as a data coordinator and data operations engineer on clinical research data for Medidata Solutions — building Python ETL pipelines, validation and QC processes for clinical studies. Meet the team.

We are engineers and data specialists, not lawyers or clinicians. Where a project needs legal, regulatory or clinical sign-off, we work alongside your advisors and domain experts.

What's included

• Extraction of drugs, conditions, adverse events, dosages and patient characteristics
• Processing pipelines for literature and case reports
• De-identification of free text and structured records
• Measured de-identification accuracy and residual-risk review
• Mapping to standard vocabularies such as ICD-10, SNOMED CT or MedDRA, where licenses permit
• Medical dataset design and expert-reviewed annotation
• Evaluation sets for medical language-processing models
• Self-hosted processing so data stays in your environment
• Documentation of every processing step

How it works

1. Agree purpose, legal basis and data-handling terms before any data is shared.
2. Pilot on a small approved sample — de-identified wherever possible.
3. Build and validate the pipeline, measuring its accuracy.
4. Deliver results with full processing documentation, then delete working copies as agreed.