Language Models for Automatic Clinical Coding in Veterinary Texts: Experimental Results

Area of Science:

• Veterinary medicine
• Artificial intelligence
• Natural language processing

Background:

• Accurate coding of veterinary diagnoses is crucial for data analysis and clinical decision-making.
• Existing automated coding systems face challenges in handling the complexity of veterinary terminology.
• Generative artificial intelligence models offer potential for improving automated medical coding.

Purpose of the Study:

• To evaluate the performance of six medium-sized generative models on automated veterinary ICD-11 coding.
• To identify the most reliable model for this specific task.

Main Methods:

• Six generative models, including Autoregressive Transformers and a Diffusion Language Model, were assessed.
• Performance was evaluated using multiple reliability metrics for ICD-11 coding accuracy.

Main Results:

• Gemma 3n demonstrated superior performance compared to the other five models.
• Gemma 3n achieved the highest reliability scores across all evaluated metrics.

Conclusions:

• Gemma 3n is a highly reliable model for automated veterinary ICD-11 coding.
• The findings suggest Gemma 3n can significantly enhance the efficiency and accuracy of veterinary record-keeping.