An Automated Framework for Calf Muscle Activity Classification: An AOA-Optimized LSTM Approach on Fused sEMG-Accelerometer Data

Accurate classification of calf muscle activity is crucial for applications in rehabilitation, clinical mobility assessment, and assistive technologies. This study presents and evaluates an automated framework using the Archimedes Optimization Algorithm (AOA) to optimize a Long Short-Term Memory (LSTM) network for classifying activity based on fused surface Electromyography (sEMG) and accelerometer data. Data were collected from eight healthy participants performing six distinct low- and high-mobility tasks, including a protocol to simulate venous stasis. After signal processing, a Correlation-based Feature Selection (CFS) method identified the most salient features for model input. The proposed AOA-optimized LSTM model achieved a high classification accuracy of 99.0% under these controlled conditions, significantly outperforming both a standard LSTM (92.1%) and a classical SVM (87.4%). The model's robustness and generalization capabilities were confirmed through 10-fold cross-validation (98.2% mean accuracy) and bootstrap analysis (84.9% mean accuracy), with the latter providing a realistic performance benchmark. These findings indicate that automating hyperparameter optimization via AOA is a promising proof-of-concept for developing high-fidelity classification models for complex biosignals. Although the offline optimization phase is computationally intensive, the resulting trained LSTM model has the potential for implementation in real-time wearable systems. This work should be viewed as a foundational feasibility study rather than a ready-for-deployment solution.