Machine Learning Approaches to Predict 24-Hour Urine Collection Results Based on Self-Reported Beverage Intake

Abstract

Objective

Repeated 24-hour urine (24HU) collections are used to evaluate risk factors for recurrence of kidney stones but are costly and burdensome. This study aimed to develop and validate machine learning models to predict 24HU volume from patients’ self-reported beverage intake and classify compliance with guidelines for preventing kidney stone recurrence.

Repeated 24-hour urine (24HU) collections are used to evaluate risk factors for recurrence of kidney stones but are costly and burdensome. This study aimed to develop and validate machine learning models to predict 24HU volume from patients’ self-reported beverage intake and classify compliance with guidelines for preventing kidney stone recurrence.

Design and Methods

Data were extracted from 2 clinical trials: trial 1 (development dataset: n = 380) and trial 2 (validation dataset: n = 142). Regression models (linear regression, regression trees, random forest, and support vector machine) were trained to predict continuous 24HU volume, while classification models (logistic regression, classification trees, random forest, and support vector machine) were trained to predict low urine volume (<2 L).

Data were extracted from 2 clinical trials: trial 1 (development dataset: n = 380) and trial 2 (validation dataset: n = 142). Regression models (linear regression, regression trees, random forest, and support vector machine) were trained to predict continuous 24HU volume, while classification models (logistic regression, classification trees, random forest, and support vector machine) were trained to predict low urine volume (<2 L).

Results

No differences were found between the development and validation datasets on demographic characteristics, 24HU volume, or self-reported beverage intake. Random forest model performed the best in predicting 24HU volume on both training and external validation datasets. Random forest also excelled at predicting high urine output in the development dataset but with overfitting risks. All classification models had high negative predictive values, reliably identifying individuals with low urine volume.

No differences were found between the development and validation datasets on demographic characteristics, 24HU volume, or self-reported beverage intake. Random forest model performed the best in predicting 24HU volume on both training and external validation datasets. Random forest also excelled at predicting high urine output in the development dataset but with overfitting risks. All classification models had high negative predictive values, reliably identifying individuals with low urine volume.

Conclusions

Machine learning models based on self-reported beverage intake and demographic characteristics can predict 24HU volume in kidney stone patients. They reliably identified patients with ≥2 L/day urine output but perform poorly for identifying those with low output volumes. Additional inputs should be considered to improve prediction and help identify patients who would benefit from targeting fluid intake for stone prevention.

Machine learning models based on self-reported beverage intake and demographic characteristics can predict 24HU volume in kidney stone patients. They reliably identified patients with ≥2 L/day urine output but perform poorly for identifying those with low output volumes. Additional inputs should be considered to improve prediction and help identify patients who would benefit from targeting fluid intake for stone prevention.