Multimodal CNN–BiLSTM for Cross-Subject Upper-Limb Motion Classification: Quantifying

In this video, we break down our latest research tackling cross-subject variability in EMG-based intent recognition. By fusing surface EMG, accelerometer, and glove data through a parameter-efficient multimodal deep learning model (1D CNNs + BiLSTM + Attention), we achieved a 77.54% average accuracy across 53 different movement classes. Designed specifically for real-world, embedded prosthetic systems, our model features: 🔹 1.29M parameters for a lightweight footprint 🔹 Less than 10 ms inference latency for real-time control 🔹 Leave-one-subject-out (LOSO) evaluation to test true real-world adaptability We also dive into our confusion matrix analysis, exploring how class imbalances and physiologically similar gestures affect deep learning performance in robotics.