Fine-tuning foundational models to code diagnoses from veterinary health records

Area of Science:

• Veterinary Informatics
• Computational Linguistics
• One Health Research

Background:

• Veterinary medical records are valuable for research but face interoperability issues due to inconsistent formats and data siloing.
• Standardized medical terminologies and clinical coding enhance record quality and facilitate data sharing across institutions and species.
• Previous Natural Language Processing (NLP) studies automated subsets of veterinary diagnosis codes using specific models like LSTM and transformers.

Purpose of the Study:

• To expand automated veterinary diagnosis coding to include all 7,739 Systemized Nomenclature of Medicine - Clinical Terms (SNOMED-CT) codes.
• To evaluate the performance of 13 diverse pre-trained language models (LMs) for automated coding using veterinary electronic health records (EHRs).
• To assess the impact of data volume and LM size/type on coding accuracy.

Main Methods:

• Fine-tuned 13 distinct pre-trained language models (LMs) on free-text clinical notes from 246,473 veterinary patient visits.
• Utilized a comprehensive set of 7,739 SNOMED-CT diagnosis codes from Colorado State University Veterinary Teaching Hospital (CSU VTH) EHRs.
• Compared the performance of various LMs, including clinical and non-clinical models, with different amounts of labeled data.

Main Results:

• The fine-tuned LMs demonstrated superior performance in automated veterinary diagnosis coding compared to previous efforts.
• Optimal accuracy was achieved using large clinical LMs fine-tuned on extensive labeled data.
• Comparable results were attainable with more limited resources and non-clinical LMs, indicating accessible methods for automated coding.

Conclusions:

• Automated coding using advanced NLP techniques significantly enhances veterinary EHR quality and interoperability.
• This approach facilitates the creation of integrated, cross-species health databases, advancing veterinary and One Health research.
• Accessible methods for automated coding improve data utility and support collaborative research efforts across institutions.