Automatic recognition of exercise types using data from wearable sensors is a key challenge in the feld of digital health and ftness. This study presents a machine learning–based strategy for classifying physical activities using data collected from ftness trackers. We employ a cost-sensitive and optimized Long Short-Term Memory (LSTM) network to better handle class imbalances and model temporal dependencies in the sensor data. During preprocessing, raw signals are cleaned, normalized, and transformed into meaningful features. The customized LSTM model is then trained to learn hidden patterns with an emphasis on minimizing misclassifcation costs. Evaluation using standard performance metrics shows that the proposed costsensitive LSTM approach signifcantly outperforms traditional methods, achieving an accuracy of 99%. Precision, recall, and F1-score also indicate excellent performance, demonstrating the method’s strong capability in accurate activity recognition. Unlike many previous studies focused on non-sequential or standard LSTM models, this research highlights the advantages of tailored recurrent architectures in enhancing the robustness and accuracy of exercise classifcation systems.