← all work
01Nov 2024 — Apr 2025AI Research Scientist · NuDiscoveries

Adaptive, Privacy-Preserving Multi-Modal Healthcare Data Pipeline

A pipeline for automated extraction, structuring, and semantic mapping of unstructured medical data using large language models — built so that longitudinal patient records can be assembled without compromising privacy.

OllamaLangchainSemantic mappingLLMs
[ pipeline / system architecture diagram ]
// approach
Ingest heterogeneous document formats — ODT, DOCX, PDF, PPTX, OCR'd JPEG, and plain text — and extract their textual content.
Apply semantic filtering to segment the meaningful clinical data out of each document.
An LLM-based chunking model classifies the text into the chunks relevant to a given query.
A second LLM extraction module converts the filtered chunks into structured JSON.
A semantic matching engine maps the extracted medical entities to OMOP-compliant fields.
// outcomes
Improves data interoperability across incompatible record formats.
Cuts the manual intervention needed to structure clinical data.
Accelerates downstream AI-driven healthcare analytics.
// related patents
FILED
System and Methods for Adaptive and Privacy-Preserving Longitudinal Structuring of Multi-Modal Healthcare Data
Patent filed
← previous
next →