Principal Investigator: Prof. Basem Shihada

Poster Presenter: Luyao Yang

Lab: Network Research Lab (NETLAB)


Oxygen Uptake Prediction Using Temporal Fusion Convolutional Network for Cardiorespiratory Fitness




Cardiorespiratory fitness (CRF) plays a crucial role in assessing human health and athletic performance. It encompasses the efficiency and capability of the cardiovascular and respiratory systems to deliver oxygen to the muscles during physical exertion. Oxygen uptake (VO2) is a widely recognized clinical and physiological indicator of CRF and exercise capacity. A common professional method to evaluate VO2 is through the utilization of the cardiopulmonary exercise test (CPET). However, its utilization is significantly constrained due to the substantial high expenses and the scarcity of trained professionals capable of conducting and interpreting CPET outcomes. In this study, we introduce an AI-based system based on a designed end-to-end Temporal Fusion Convolutional Network (TFCN) to predict the VO2 dynamics. Our system employs physiological parameters measured by accessible wearable sensors, such as heart rate (HR), heart rate reserve (HRR), minute ventilation (VE), tidal volume (VT), and breathing frequency (BF), to accurately predict VO2 dynamics. These variables were derived from CPET data collected from a diverse group of 58 adults using a COSMED system. The group consisted of 41 males and 17 females, with 38 individuals classified as healthy and 20 as smokers. The average age of the participants was 30±15 years, with an average height of 1.75±0.2m, weight of 90±30kg, and VO2 max of 42±6 L/min. Out of the total dataset, 40 data samples were used for training the TFCN model, while the remaining 18 data samples were reserved for testing and evaluating the model's performance. Each participant was subjected to an incremental exercise protocol, facilitating a comprehensive exploration of VO2 dynamics. In comparison to other models such as LSTM (RMSE: 0.3 L/min, R2: 0.69) and TCN (RMSE: 0.27 L/min, R2: 0.8), our TFCN model showcases the state-of-art performance, achieving an RMSE of 0.23 L/min and an R2 of 0.92. Given the exceptional and dependable performance exhibited by our proposed system, we are now motivated to validate this affordable wearable sensor system within a larger population in Saudi Arabia.