This paper proposes a real-time framework to estimate the energy expenditure (EE) in gym exercises by combining the information provided by IoT wearables and machine learning (ML). The classical techniques, like indirect calorimetry are both expensive and immobile whereas only the heart rate model mostly fails to work in dynamic situations. To overcome these drawbacks, we are going to rely on a multimodal dataset featuring a combination of physiological, personal, and environmental measurements and captured with a Zephyr BioHarness 3 device. We learn and train supervised models of learning, such as recurrent (LSTM, CNN-LSTM) and two new hybrids: an LSTM-LightGBM ensemble and a CNN-LSTM-LightGBM cascade. The proposed system identifies gym activities, and at the same time, the estimated values of EE are accurate and real time, which is a viable alternative to the lab-based methods. We have shown that our experimental assessment provides a high level of predictive power with the Random Forest model, which gives us an R 2 = 0.94 and RMSE = 0.39 MET, and with the LSTMLightGBM ensemble, which gives us R 2 = 0.80 and RMSE = 0.71 MET.Key contributions include a subject-independent training methodology for robust generalization and a feedback mechanism for monitoring athletic performance. Experimental results validate the precision and responsiveness of the framework, demonstrating its potential in both sports science and clinical healthcare. By integrating IoT sensing with ML, this work advances adaptive data-driven approaches to optimize training and recovery.
Nazeer et al. (Fri,) studied this question.