Nonexercise machine learning models for maximal oxygen uptake prediction in national population surveys

Area of Science:

• Cardiovascular Health
• Biostatistics
• Machine Learning

Background:

• Nonexercise algorithms estimate cardiorespiratory fitness (CRF) cost-effectively.
• Existing algorithms have limitations in generalizability and predictive accuracy.
• Improving CRF estimation is crucial for public health and clinical practice.

Purpose of the Study:

• To enhance nonexercise algorithms for CRF estimation using machine learning (ML).
• To leverage US national population survey data for improved predictive models.
• To develop more accurate VO2 max estimation methods.

Main Methods:

• Utilized data from the National Health and Nutrition Examination Survey (NHANES) (1999-2004).
• Applied multiple ML algorithms, including LightGBM, to predict maximal oxygen uptake (VO2 max).
• Developed a parsimonious model (interview/examination data) and an extended model (including DEXA and lab tests).

Main Results:

• The LightGBM algorithm demonstrated superior performance.
• The parsimonious ML model reduced VO2 max estimation error by 15% (RMSE: 8.51).
• The extended ML model reduced error by 12% (RMSE: 8.26), showing significant improvements over existing methods.

Conclusions:

• ML integration with national data offers a novel approach to CRF estimation.
• The developed nonexercise models provide improved accuracy for VO2 max prediction.
• These findings support better cardiovascular disease risk classification and clinical decision-making.