Automatic inference of ICD-10 codes from German ophthalmologic physicians' letters using natural language processing

Area of Science:

• Medical Informatics
• Natural Language Processing
• Ophthalmology

Background:

• Physicians' letters are rich diagnostic sources but often lack standardized codes.
• Medical registries frequently require diagnosis codes like ICD-10 for data analysis.
• Bridging the gap between free-text clinical notes and structured coding is crucial for data utilization.

Purpose of the Study:

• To develop and evaluate an algorithm for inferring ICD-10 codes from German ophthalmologic physicians' letters.
• To assess the accuracy of the algorithm across multiple clinical settings.
• To determine the feasibility of using this method for improving registry data quality.

Main Methods:

• An algorithm based on the nearest-neighbor method and a comprehensive ICD-10 thesaurus was developed.
• The thesaurus was embedded into a Word2Vec space using anonymized physician reports.
• The algorithm's performance was evaluated on diagnoses from 100 letters each at three German eye hospitals.

Main Results:

• The algorithm successfully inferred ICD-10 codes from 2806 natural language diagnoses across three hospitals.
• Accuracy reached 98% for fully correct codes and 99% for superordinate concepts in the first hospital.
• Accuracy varied across hospitals, with 69%-91% correct codes, suggesting domain-specific embedding benefits.

Conclusions:

• The developed algorithm effectively infers ICD-10 codes from German clinical text, particularly when trained on local data.
• This method offers a viable solution for enhancing medical registry data quality.
• The approach is adaptable to other languages and medical specialties.