Estimating proximity to muscular failure using surface EMG and deep learning

: Understanding how close an individual is to muscular failure during exercise can be used to personalize resistance training dynamically. We propose a deep learning approach to estimate proximity to failure in real-time from surface electromyography (sEMG) signals, supported by a novel dataset of 192 recordings collected from 12 participants performing isometric biceps brachii holds to failure. The recorded sEMG signals were preprocessed and converted into spectrogram windows using the Short-Time Fourier Transform. A continuous Proximity to Failure Index (PFI), defined as the elapsed percentage of the hold duration, was assigned to each window based on its corresponding position in time. Several deep learning models, including a multilayer perceptron, a Transformer, and recurrent neural networks, were trained to predict PFI values from the spectrograms and compared against linear and support vector regression baselines. Model performance was evaluated using leave-one-out cross-validation. All deep learning models outperformed the baselines, with the long short-term memory network achieving the lowest mean squared error (49.44±18.34) across participants. This work demonstrates that proximity to muscular failure can be estimated from sEMG signals during isometric holds, offering a basis for developing real-time biofeedback systems that adapt resistance training according to electromyographic activity.