A Generalized Machine Learning Model for Identifying Congenital Heart Defects (CHDs) Using ICD Codes.

Area of Science:

• Medical Informatics
• Cardiology
• Machine Learning

Background:

• International Classification of Diseases (ICD) codes have high false-positive rates for identifying congenital heart defect (CHD) cases.
• Improved accuracy in CHD identification is crucial for effective public health surveillance.

Purpose of the Study:

• To evaluate the effectiveness of machine learning (ML) algorithms in enhancing the accuracy of CHD case identification.
• To assess if ML can improve upon traditional ICD code-based case selection.

Main Methods:

• Applied traditional ML methods to four encounter-level datasets (2010-2019) for 3334 patients with validated CHD diagnoses and ICD codes.
• Utilized a 5-fold cross-validation approach to identify key features for CHD classification.
• Explored various training and testing combinations to optimize CHD classification accuracy.

Main Results:

• Initial CHD ICD positive predictive values (PPVs) varied significantly by site (53.2%–84.0%).
• The ML algorithm achieved a high PPV of 95% for the combined dataset, with a false-negative (FN) rate of 33% when prioritizing PPV.
• XGBoost effectively reduced 2105 Clinical Classification Software (CCS) features to 137, distinguishing true-positive CHD cases from false positives.

Conclusions:

• Integrating ML algorithms after ICD code selection substantially improves the accuracy of identifying true-positive CHD cases.
• ML offers a promising approach to refine CHD surveillance by mitigating the limitations of ICD codes.