Machine learning in public health informatics: Evidence that complex sampling structures may not be needed for...

Area of Science:

• Health Informatics
• Biostatistics
• Machine Learning

Background:

• National health surveys often have imbalanced outcome data, posing challenges for predictive modeling.
• Traditional statistical models may not optimally handle complex survey designs and imbalanced data.
• Machine learning offers alternative approaches for analyzing large-scale health survey data.

Purpose of the Study:

• To assess the predictive performance of machine learning models for imbalanced outcomes in national health surveys.
• To compare machine learning models against traditional logistic regression incorporating complex sampling designs.
• To evaluate models without the use of survey sampling weights.

Main Methods:

• Utilized US National Health and Nutrition Examination Survey (USNHANES) data.
• Compared support vector machine, random forest, LASSO regression, and deep neural networks.
• Employed oversampling, undersampling, and combined resampling techniques for class imbalance.
• Benchmarked against logistic regression with complex sampling design integration.

Main Results:

• Machine learning models demonstrated comparable balanced accuracy (0.72-0.76) to traditional methods.
• Support vector machine and neural networks excelled in sensitivity (0.79-0.83).
• Random forest achieved the highest specificity (0.86-0.96).
• Area Under the Curve (PR-AUC) and Brier scores indicated varying model performance.

Conclusions:

• Machine learning models offer similar predictive power to established methods for osteoarthritis prediction using USNHANES data.
• The evaluated machine learning models can effectively predict imbalanced outcomes without sampling weights.
• These findings support the use of machine learning in analyzing complex national health survey data.