Improving MedDRA/J Coding Accuracy with a Fine-Tuned Text Embedding Model.

Area of Science:

• Pharmacovigilance
• Natural Language Processing
• Medical Coding

Background:

• MedDRA (Medical Dictionary for Regulatory Activities) coding in Japan presents challenges due to discrepancies between source expressions and dictionary terms.
• Accurate adverse event (AE) reporting is crucial for patient safety and regulatory compliance.

Purpose of the Study:

• To enhance the accuracy and efficiency of MedDRA coding in Japan.
• To develop and evaluate a fine-tuned text embedding model using in-house AE data.

Main Methods:

• A text embedding model was fine-tuned on a dataset of 50,000 in-house adverse event entries.
• The performance of the fine-tuned model was compared against baseline methods for MedDRA term ranking and recall.

Main Results:

• The fine-tuned model demonstrated significant improvements in MedDRA term ranking and recall.
• Achieved nDCG@20 of 76.2% and Recall@20 of 90.8% with 50K in-house entries.
• Outperformed baseline approaches in accurately mapping source expressions to MedDRA terms.

Conclusions:

• Fine-tuning text embedding models on real-world AE data is a viable strategy to improve MedDRA/J coding.
• This advanced approach can significantly benefit pharmacovigilance activities by increasing search accuracy and efficiency.
• The methodology offers a scalable solution for addressing coding discrepancies in regulatory settings.