Prediction of 24-Hour Urinary Sodium Excretion Using Machine-Learning Algorithms

Area of Science:

• Nutritional epidemiology
• Biostatistics
• Machine learning applications in health

Background:

• Accurate quantification of dietary sodium intake is challenging using self-reported methods.
• Self-reported dietary assessments often suffer from measurement error.

Purpose of the Study:

• To apply machine learning (ML) algorithms for predicting 24-hour urinary sodium excretion.
• To assess the performance of ML in estimating sodium intake compared to traditional methods.

Main Methods:

• Utilized an ensemble ML approach with 36 characteristics to predict 24-hour urinary sodium excretion in 3454 participants from NHS, NHS-II, and HPFS.
• Externally validated ML algorithms using data from the Trial of Hypertension Prevention I (TOHP-I).
• Applied ML to estimate hazard ratios of incident hypertension associated with predicted sodium intake in 167,920 adults over 30 years.

Main Results:

• ML demonstrated superior prediction and calibration of 24-hour urinary sodium excretion compared to food frequency questionnaires (Spearman correlation 0.51 vs. 0.19).
• ML prediction accuracy was moderately improved for energy-adjusted sodium excretion.
• ML-predicted sodium showed a modestly stronger association with incident hypertension in NHS-II, but not in NHS or HPFS.

Conclusions:

• Developed ML algorithms significantly enhance the prediction of absolute 24-hour urinary sodium excretion.
• The ML approach shows potential for generalizable sodium intake prediction.
• ML does not substantially mitigate bias from measurement error in disease association studies.