A machine learning model for predicting congenital heart defects from administrative data.

Area of Science:

• Medical Informatics
• Public Health
• Machine Learning

Background:

• International Classification of Diseases (ICD) codes are commonly used for identifying congenital heart defects (CHD) in administrative data.
• However, ICD codes may inaccurately identify true positive (TP) individuals with CHD, potentially weakening surveillance efforts.

Purpose of the Study:

• To enhance CHD surveillance by accurately identifying TP CHD cases in administrative records using machine learning (ML) algorithms.
• To identify key features that predict TP CHD individuals within large administrative datasets.

Main Methods:

• Applied traditional ML models (Logistic Regression, Gaussian Naive Bayes, Random Forest, XGBoost) to a validated dataset of 779 patients.
• Utilized encounter-level data including ICD-9-CM and CPT codes from 2011-2013 across four US sites.
• Employed five-fold cross-validation to identify overlapping important features and compared model performance using metrics like PPV and F1-score.

Main Results:

• Expert clinician validation of ICD-9-CM CHD codes yielded a baseline positive predictive value (PPV) of 76.5%.
• ML feature selection reduced 7138 features to 10, significantly improving prediction of TP CHD cases.
• eXtreme Gradient Boosting (XGBoost) demonstrated superior performance, achieving a median PPV of 0.94 (95% CI: 0.94, 0.95) and F1-score of 0.84 (95% CI: 0.76, 0.91).

Conclusions:

• ML algorithms substantially improve the accuracy of identifying TP CHD cases compared to using ICD codes alone.
• This ML approach enhances the generalizability of findings from large datasets to the CHD patient population.
• Improved identification accuracy strengthens public health surveillance for congenital heart defects.