Introduction: Water is the most abundant molecule in the human body and supports major functions of life. Hydration represents the balance between water intake and loss, and studies have shown that higher levels of hydration are associated with lower mortality risks. However, the United Nations estimates that nearly one quarter of the world’s population does not have access to sufficient drinking water, leading to more than 3.5 millions deaths each year. Even minor levels of water loss can lead to impacts on the renal system (e.g., kidney stones), the urinary system (e.g., urinary tract cancers), and diminished physical and mental performance. Given the criticality of water in biological systems and the complexity of its maintenance, a number of reference measures have recently been proposed to determine hydration level including body mass, urine and blood concentration, and self-reported thirst, among others. However, existing reference measures are often in conflict due to varying levels of time scales, targets, and levels of accuracy. The goal of the current study is to combine reference measurements in a way to comprehensively assess body hydration with high accuracy. Such an approach may be useful in real-time assessments of hydration status, while also serving as a standard against which new measures may be validated.
Materials and
Methods: A sample size of N=160 human subjects was estimated to develop a binary hydration classifier based on reference data using a randomized, cross-over design. Subjects were asked to alter their fluid consumption on consecutive days by either following a hypohydration diet (17.5 ml/kg) or a hyperhydration diet (52.5 ml/kg) for 24 hours prior to in-lab appointments. Presentation of hyper- and hypohydration days were counterbalanced. Subjects kept track of fluid types and amounts consumed, and foods eaten in an electronic log. Subjects were instructed to maintain their typical daily food intake and minimize caffeine and alcohol intake. Upon arrival, experimenters reviewed dietary fluid logs to ensure compliance. Subjects were then asked to void completely and rest for one hour. Subjects were then asked to void again and provide a ~2.5 ml urine sample for assessment of urine osmolality and urine specific gravity (USG). Subjects also provided an ~70 µl blood sample via finger lancet for assessment of plasma osmolality, a body mass measurement, and self-reported thirst using both a visual analog scale (VAS) and a categorical scale. A subset of subjects also provided an additional 10 µL blood sample for hemoglobin (Hb) quantification. Subjects were then asked to commence the opposite diet, and returned to the lab for repeat measurements the following day. All analyses were performed in Python using Pandas, Matplotlib, Numpy, Scipy, Seaborn, and Sci-kit learn packages. Paired t-tests were performed to compare metrics between hypohydration and hyperhydration states. Multicollinearity analyses were performed to define features for hydration classification.
Results, Conclusions, and Discussions:
Results: A total of 162 subjects were recruited for this study (81 male, 81 female), with an average age 37.3 ± 10.1 years (range 22 to 70 years). The hyperhydration diet was associated with a significant increase in reported fluid consumption (p = 3.51x10-87) and body mass (p = 1.20 x 10-5). The hypohydration diet was associated with a significant increase in plasma osmolality (p = 2.9 x 10-10), urine osmolality (p = 1.05 x 10-34), USG (p = 3.84 x 10-33), and thirst using both the categorical scale (p = 2.04 x 10-34) and the VAS (p = 6.55 x 10-34). No changes in Hb were observed based on changes to fluid intake (p = 0.71). Correlations between all reference metrics and demographic variables are shown below, with delta values indicative of hyperhydration - hypohydration, and significant correlations shown with an asterisk. Several metrics were highly correlated, including USG and urine osmolality (r=0.97), and the thirst scales (r=0.88). The urine metrics correlated significantly with fluids consumed (r=0.21), thirst (r=0.32), and demographic variables, and plasma osmolality correlated significantly with fluids consumed (r=0.4). Highly correlated features were removed from model development, including body mass index (BMI), urine osmolality, and thirst from the categorical scale. A logistic regression model was trained using Scikit-learn with a 75:25 training:test split and incorporated demographic features, including: sex, age, and height; and reference features including: weight, plasma osmolality, USG, and thirst from the VAS. Features were used as predictors of hypohydration or hyperhydration status. The model classified hydration status at an F1 level of 87%.
Conclusion: The current study demonstrated a highly accurate and explainable model of human hydration status using an ensemble of reference features. Such an approach is useful in providing a discrete assessment of hydration status that may be useful for individuals at risk of dehydration, or who may be undergoing clinical care. This approach is also valuable for assessing the accuracy and validity of emerging point of care devices and algorithms that assess hydration status using wearable devices.
