Finger Force Prediction from Spinal Signals: Machine Learning Pipeline for the Neural Drive

Estimating muscle force production from neural drive is essential for advanced human-machine interfaces and understanding motor control. Although motor unit (MU) spike trains derived from high-density surface electromyography (HD- sEMG) have been used in gesture recognition and force estimation, the role of neural drive feature granularity (i.e. level of detail in its representation) in force regression remains understudied. In this study, we recorded HDsEMG of the anterior forearm from 25 participants performing isometric finger flexion and compare three sets of neural drive-derived features of increasing granularity: discharge rates (DRs) of ten representative MUs, DRs from two cumulative spike trains (CSTs) from 5 early and 5 late recruited MUs, and DRs from a single CST from all 10 MUs. These features were used to train regression models including linear models, XGBoost, multilayer perceptrons (MLPs), and recurrent neural networks (RNN, GRU, LSTM). Our results show that higher granularity of neural drive-derived features improves prediction, particularly in LSTM models, which achieved coefficient of determination values up to 0.944. Unlike prior subject- and task-specific approaches, the proposed pipeline generalises across participants and digits. These findings demonstrate that MU-level detail is crucial for accurate force regression and offer a foundation for generalisable MU-based neural interfaces.

Recommended citation: R. Mio, J. Bodenschlägel and A. A. Faisal, “Finger Force Prediction from Spinal Signals: Machine Learning Pipeline for the Neural Drive,” 2025 International Conference on Neural Engineering (NER), San Diego, CA, USA, 2025, pp. 1194-1199.
Download Paper