Personalized prediction of optimal water intake in adult population by blended use of machine learning and clinical data

Abstract

Growing evidence suggests that sustained concentrated urine contributes to chronic metabolic and kidney diseases. Recent results indicate that a daily urinary concentration of 500 mOsm/kg reflects optimal hydration. This study aims at providing personalized advice for daily water intake considering personal intrinsic (age, sex, height, weight) and extrinsic (food and fluid intakes) characteristics to achieve a target urine osmolality (UOsm) of 500 mOsm/kg using machine learning and optimization algorithms. Data from clinical trials on hydration (four randomized and three non-randomized trials) were analyzed. Several machine learning methods were tested to predict UOsm. The predictive performance of the developed algorithm was evaluated against current dietary guidelines. Features linked to urine production and fluid consumption were listed among the most important features with relative importance values ranging from 0.10 to 0.95. XGBoost appeared the most performing approach (Mean Absolute Error (MAE) = 124.99) to predict UOsm. The developed algorithm exhibited the highest overall correct classification rate (85.5%) versus that of dietary guidelines (77.8%). This machine learning application provides personalized advice for daily water intake to achieve optimal hydration and may be considered as a primary prevention tool to counteract the increased incidence of chronic metabolic and kidney diseases.