Czech medical coding assistant based on transformer networks.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Manual clinical coding of medical reports using the International Classification of Diseases (ICD) in the Czech Republic is prone to errors and costly due to the human factor.
• Existing processes require extensive coder training and significant effort per report, leading to inaccuracies.

Purpose of the Study:

• To develop and implement an AI-powered system assisting clinical coders by automatically predicting ICD diagnosis codes.
• To enhance the efficiency and accuracy of medical report coding through automated predictions presented for coder review.

Main Methods:

• Development of small, transformer-based models for ICD classification in the Czech language, focusing on minimal memory consumption, generality, portability, and sustainability.
• Introduction of a novel "Four-headed" transformer model with multiple classification heads, designed to economize memory and learning time while maintaining performance.
• Evaluation of models on Czech datasets for main and all diagnosis classification tasks, and comparison with state-of-the-art English models using the Mimic IV dataset.

Main Results:

• The proposed transformer-based models achieve comparable or superior results to individual models, with the "Four-headed" model offering significant savings in memory and training time.
• The developed models demonstrate comparable performance to state-of-the-art English models despite being substantially smaller.
• The models are optimized for swift utilization on standard hospital computer systems and allow for easy retraining with new data.

Conclusions:

• The developed AI system effectively assists clinical coders by automating ICD code prediction, thereby improving efficiency and accuracy in medical report processing.
• The proposed transformer models, particularly the "Four-headed" architecture, offer a computationally efficient and effective solution for Czech clinical coding.
• The study highlights the potential of smaller, specialized models for accurate and sustainable medical code prediction in resource-constrained environments.