Selecting relevant features from the electronic health record for clinical code prediction | JoVE Visualize

Area of Science:

Health Informatics
Medical Data Analysis
Artificial Intelligence in Healthcare

Background:

Electronic health records (EHRs) contain vast, heterogeneous data sources crucial for automated clinical coding.
Information overlap and varying quality across EHR sources complicate accurate diagnosis and procedure code assignment.
Effective feature selection is necessary to create a dense, high-quality data representation for improved coding.

Purpose of the Study:

To introduce and evaluate coverage-based feature selection methods for clinical code prediction.
To compare confidence and information gain approaches for integrating diverse EHR data sources.
To assess the performance of these methods across multiple medical specialties for ICD-9-CM and ICD-10-CM code prediction.

Main Methods:

Developed and compared coverage-based feature selection techniques, specifically confidence and information gain.
Evaluated methods on seven medical specialties for ICD-9-CM code prediction (Antwerp University Hospital and MIMIC-III) and two for ICD-10-CM.
Compared performance against a baseline feature selection (frequency threshold) and individual data sources.

Main Results:

Confidence coverage consistently improved F-measure for diagnosis code prediction (49.83% average) compared to baseline (44.25%) and best standalone source (44.41%).
The method created concise, interpretable patient stay representations, independent of frameworks like UMLS.
Utilizing multiple heterogeneous EHR sources with confidence coverage enhanced prediction accuracy while reducing features and processing time.

Conclusions:

Coverage-based feature selection, particularly confidence coverage, offers a robust method for integrating diverse EHR data for clinical code prediction.
This approach effectively addresses information overlap and quality issues, leading to significant improvements in coding accuracy.
Confidence coverage provides a flexible and interpretable alternative to traditional methods, enhancing the efficiency of automated clinical coding systems.