ECG-Based Energy Expenditure Prediction Using BiLSTM With Improved Snow Ablation Optimizer

Accurately estimating energy expenditure (EE) is essential for optimizing motion control strategies in human-in-the-loop (HIL) systems within the field of human-machine interface (HMI). This study aims to propose a method for accurately estimating EE under various movement conditions using a low-cost wearable electrocardiogram (ECG) device and to analyze the impact of various factors on the estimation results. In this study, we introduce an improved snow ablation optimizer (ISAO) to optimize the hyperparameters and input time series window length in a BiLSTM network for predicting EE under diverse motion conditions. ISAO adaptively adjusts the search strategy during different optimization stages by analyzing fitness changes throughout the iterative process. Additionally, the optimized input data window length considers both high-frequency ECG signals and low-frequency EE data collected using the breath-by-breath method, thus avoiding errors from empirical selection. After offline model training, we evaluated the model using test sets comprising multiple time intervals generated by the optimized window length, simulating real online testing scenarios. Experimental results demonstrate that the proposed ISAO-BiLSTM network structure can accurately predict EE across various motion paradigms (RMSE: 56 W; ${R}^{{2}}$ : 0.93). This study verifies that the ECG-based deep learning framework has more accurate EE predictions across various motion paradigms compared to those based on heart rate (HR) signals. Additionally, the optimized input data window length is expected to be valuable for real-time EE predictions in HIL HMI applications, as current closed-loop optimization strategies primarily rely on outputting an EE value within a specific time window.