Using a Large Open Clinical Corpus for Improved ICD-10 Diagnosis Coding.

Area of Science:

• Natural Language Processing
• Deep Learning
• Medical Informatics

Background:

• Advances in NLP and deep learning make AI-assisted medical coding more feasible.
• Efficient ICD-10 diagnosis coding of discharge summaries is crucial for healthcare data management.

Purpose of the Study:

• To present novel datasets for training AI models for medical coding.
• To develop and evaluate a BERT-based language model for assigning ICD-10 codes to Swedish discharge summaries.
• To demonstrate the utility of these datasets and models in a practical coding support tool.

Main Methods:

• Development and curation of specialized datasets for medical coding tasks.
• Training a BERT-based language model on Swedish discharge summaries.
• Evaluation of the model's performance in assigning ICD-10 codes.
• Simulation of a coding support tool recommending codes to reduce coder workload.

Main Results:

• A BERT-based language model was successfully trained using the presented dataset.
• The model demonstrated consistent performance in assigning accurate ICD-10 codes to Swedish discharge summaries.
• The developed model can be integrated into a coding support system to enhance coder efficiency.

Conclusions:

• The presented datasets are valuable resources for developing AI in medical coding.
• AI models, like the BERT-based one, can significantly assist medical coders by reducing their workload.
• The de-identified and pseudonymised dataset is available for academic research to foster further development.