Contrastive diagnostic embedding (CDE) model for automated coding - A case study using emergency department...

Area of Science:

• Medical Informatics
• Computational Linguistics
• Machine Learning

Background:

• Manual medical billing code assignment is costly, time-consuming, and prone to errors.
• Existing automated billing code strategies struggle with the large, dynamic ICD hierarchical structure.
• Need for robust and generalizable automated solutions for medical billing and clinical data analysis.

Purpose of the Study:

• To develop a weakly supervised contrastive learning strategy for accurate automated medical billing code assignment.
• To capture fine semantic variations between diagnosis codes, including rare and new codes.
• To enhance the generalizability and robustness of automated billing systems.

Main Methods:

• Proposed Contrastive Diagnosis Embedding (CDE) using a two-phase contrastive training approach.
• Leveraged contrastive learning to create a semantic embedding space incorporating ICD-10 hierarchy.
• Employed a weakly supervised retrieval scheme capable of handling the entire 70K ICD-10 code set.

Main Results:

• CDE model outperformed existing methods (string-matching, ClinicalBERT, CAML, PLM-ICD) on multiple test datasets.
• Demonstrated accurate prediction of rare and newly introduced diagnosis codes.
• Ablation studies confirmed the effectiveness of each training phase in improving performance.

Conclusions:

• Weakly supervised contrastive learning (CDE) overcomes limitations of rule-based and supervised models in generalization and robustness.
• The proposed model enables flexible and accurate automated billing code assignment.
• Enhances efficiency in value-based care environments through improved practice convergence.