Empowering Prediction of Resting Energy Expenditure in Free-Living Settings by AI Tools: Application of a...

Traditional predictive equations derived from regression analyses exhibit varying degrees of accuracy in estimating resting energy expenditure (REE). AI models can increase the predictability of such equations, even for population-specific ones. This work aimed to improve the prediction of REE in a dataset of Saudi population-specific equations using suitable AI tools.

The dataset from the previously published Saudi population-specific equation by Almajwal and Abulmeaty (AA) in 2019 was used to develop an artificial neural network (ANN)-based version to better predict REE in the adult population. Anthropometric and body composition parameters were used as proposed features. The proposed hybrid prediction model underwent an extensive two-stage, iterative training process. First, the Extreme Gradient Boosting (XGBoost) model is used to compute feature importance scores. Then, the most prominent features were identified and incorporated into the ANN model. These significant features were used to train the ANN model to capture nonlinear correlations among them and make accurate predictions. Subsequently, XGBoost and Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) are used for their ability to provide a multi-layer abstraction of complex input data.

A total of 423 participants (208 male, 215 female) were divided into three non-overlapping sets: training (295, 70%), validation (64, 15%), and testing (64, 15%). The ANN model, combined with XGBoost, helped us to develop two equations: AA_ANN1= 2.47 × BMI + 11.9 × AdjBW + 962.5 and AA_ANN2 = 4.29 × age + 9.4 × fat mass + 15.71 × FFMI + 1289.3, where BMI is Body Mass Index (kg/m²), AdjBW is Adjusted Body Weight (kg), and FFMI is Fat Free Mass Index (kg/m²). The AA_ANN1 presented a Root Mean Square Error (RMSE) of 215 and an accuracy of 66.2%, whereas AA_ANN2 presented a lower RMSE of 193 and a higher accuracy of 71.4%. The ANN model was trained on the top 10 features ranked by XGBoost, achieving an average accuracy of 90.2%.

The two new predictive equations, developed using an ANN combined with XGBoost, significantly improved REE prediction accuracy to 90.2%, achieved only with the full ANN model. Future external validation in an independent cohort is essential before clinical application of these equations.