An Advanced LSTM Model to Enhance Running Economy

Abstract

Running Economy (RE) is an important physiological measure for endurance athletes, particularly distance runners. It is defined as the energy demand for a specific velocity of submaximal running and depends on biomechanical, metabolic, cardiorespiratory, and neuromuscular factors. In this work, we introduce a novel approach for predicting running economy in amateur runners through the analysis and comparison of three neural network architectures: a proposed Long Short-Term Memory (LSTM) model, an alternative LSTM model, and a Recurrent Neural Network (RNN). Each model incorporates physiological and biomechanical metrics, including pace, heart rate, power, cadence, ground contact time, and stride length, collected with a wearable device. By analyzing these temporal sequences, the models estimate running efficiency using initial test data as a baseline. Their performance was evaluated using statistical metrics, including Mean Squared Error (MSE), Mean Absolute Error (MAE), and the coefficient of determination (R2). The best results obtained from the proposed LSTM Model were a test MSE of 0.036995, a test MAE of 0.102474, and a test R2 of 0.986986, demonstrating superior predictive capacity compared to the other architectures. These results highlight the potential of integrating artificial intelligence with wearable technology to provide accessible and personalized training tools for coaches and amateur athletes.