(P03-062-24) Prediction of Resting Energy Expenditure in Individuals With Obstructive Sleep Apnea Using Machine Learning

Objectives: Individuals with obstructive sleep apnea condition need to control nutrient intake when receiving clinical care. Resting Energy Expenditure (REE) is estimated using predictive energy equations to guide the nutrient support in such cases. However current calculation methods only reach up to 45% of accuracy when compared with true REE values obtained by handheld indirect calorimetry devices. This study aims to develop an optimized model for better performance in REE prediction.

Methods: Dataset was obtained from a single office of a private sleep medicine practice in Houston, Texas.  The dataset contains data from 160 subjects with 13 attributes (height, weight, sex, age, etc.). Several regression machine learning models, namely Linear regression, Gradient Boosting Regression, Decision tree, Random Forest, K-Nearest Neighbors and Neural network models were developed and their predictions compared to REE measured by indirect calorimetry.

Results: Among the models developed, Gradient Boosting Regression had the highest performance with 65.6% of predicted values falling within the band of acceptable agreement (10% of real value), followed by Random Forest (62.5%) and Linear regression as the lowest precision with only 37.5% of predictions within the band of acceptable agreement.

Conclusions: We developed six machine learning models for the prediction of REE in patients with Sleep Apnea. The results show that Gradient Boosting Regression has better predictive accuracy than established REE equations. Further validation may provide support in confirming and optimizing this approach.

Funding Sources: This work was supported by the Department of Computer Science at Texas Southern University.