Ontology- and LLM-based data harmonization for federated learning in healthcare

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Bioinformatics

Background:

• Semantic heterogeneity in electronic health records (EHRs) hinders scalable and privacy-preserving healthcare analytics.
• Federated learning (FL) offers a solution for collaborative modeling without raw data sharing, but requires consistent, ontology-aligned data.
• Current data harmonization methods are often manual and time-consuming.

Purpose of the Study:

• To develop and evaluate an ontology- and large language model (LLM)-based data harmonization approach.
• To support secure, interoperable federated learning (FL) workflows in healthcare.
• To transform ontology-based harmonization into a reusable and configurable workflow.

Main Methods:

• A two-step pipeline was proposed: 1) Candidate concept retrieval using embedding-based similarity or ontology cross-references. 2) LLM-based semantic validation of candidates against a target ontology.
• The approach is ontology-agnostic and demonstrated with mappings to MONDO and HPO.
• Final mappings were evaluated against human expert assessments.

Main Results:

• Expert-LLM agreement reached up to 92% across two clinical datasets.
• Overall performance ranged from 78% to 91%, depending on the candidate-generation strategy.
• LLM-based validation significantly improved precision, while retrieval strategies enhanced recall, outperforming retrieval alone.

Conclusions:

• The proposed pipeline automates ontology-based harmonization, making it suitable for federated healthcare research.
• Combining high-recall retrieval with LLM semantic adjudication enables scalable, privacy-preserving conversion of heterogeneous clinical text.
• This approach facilitates standardized data representations across domains for improved healthcare analytics.