Dynamic oxygen uptake (VO₂) reflects moment-to-moment changes in oxygen consumption during exercise and underpins training design, performance enhancement, and clinical decision-making. We tackled two key obstacles—the limited fusion of heterogeneous sensor data and inadequate modeling of long-range temporal patterns—by integrating wearable accelerometer and heart-rate streams with a convolutional neural network–LSTM (CNN-LSTM) architecture and optional attention modules. Physiological signals and VO₂ were recorded from 21 adults through resting assessment and cardiopulmonary exercise testing. The results showed that pairing accelerometer with heart-rate inputs improves prediction compared with considering the heart rate alone. The baseline CNN-LSTM reached R² = 0.946, outperforming a plain LSTM (R² = 0.926) thanks to stronger local spatio-temporal feature extraction. Introducing a spatial attention mechanism raised accuracy further (R² = 0.962), whereas temporal attention reduced it (R² = 0.930), indicating that attention success depends on how well the attended features align with exercise dynamics. Stacking both attentions (spatio-temporal) yielded R² = 0.960, slightly below the value for spatial attention alone, implying that added complexity does not guarantee better performance. Across all models, prediction errors grew during high-intensity bouts, highlighting a bottleneck in capturing non-linear physiological responses under heavy load. These findings inform architecture selection for wearable metabolic monitoring and clarify when attention mechanisms add value.