Automated ICD coding via unsupervised knowledge integration (UNITE).

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Accurate International Classification of Diseases (ICD) coding is crucial for medical billing and electronic medical record (EMR) research.
• Supervised methods for automatic ICD coding face challenges with data bias and portability across different EMR systems.

Purpose of the Study:

• To develop an unsupervised algorithm for automatic ICD code assignment from clinical notes.
• To evaluate the performance and portability of the developed algorithm across diverse EMR systems.

Main Methods:

• Developed the unsupervised knowledge integration (UNITE) algorithm using semantic relevance assessment of clinical narrative notes.
• Validated UNITE on ICD-coded data for 6 diseases from Partners HealthCare (PHS) Biobank and Medical Information Mart for Intensive Care (MIMIC-III).
• Compared UNITE against penalized logistic regression, topic modeling, and neural network models, assessing performance and cross-EMR portability.

Main Results:

• UNITE achieved high average AUC scores (0.91 at PHS, 0.92 at MIMIC) for 6 diseases, comparable to supervised models.
• UNITE demonstrated substantially better performance than topic models.
• The algorithm exhibited consistent and superior portability across different EMR systems compared to other evaluated methods.

Conclusions:

• The UNITE algorithm accurately assigns ICD codes in EMRs without human labor, offering advantages over existing machine learning approaches.
• UNITE provides stable performance and high portability across EMRs from different institutions, addressing key limitations of supervised methods.