Multimodal fusion of optimized LSTM for estimating dynamic oxygen uptake based on consumer wearable devices

Abstract

Accurate and real-time monitoring of oxygen consumption dynamics during exercise is essential for optimizing training intensity, duration, and recovery strategies. However, existing methods often compromise between computational efficiency and predictive accuracy, either by employing large models for feature extraction or by oversimplifying input variables, which limits their applicability to consumer wearable devices requiring real-time and energy-efficient processing. To address these challenges, we propose a novel framework that integrates Multimodal Feature Fusion and Particle Swarm Optimization (PSO) based on wearable device data. The model utilizes Long Short-Term Memory (LSTM) networks to extract temporal features from anthropometric, physiological, and kinematic indicators, while PSO optimizes feature weights to enhance prediction accuracy. The proposed framework therefore presents a 14.3% reduction in root mean squared error and 24% reduction in model size for dynamic oxygen uptake (VO2) prediction and robust performance across various moderate intensity running speeds. This multimodal fusion approach provides an effective and efficient solution for VO2 prediction, facilitating real-time deployment on consumer wearable devices.