Abstract
Previous literature shows that deep learning is an effective tool to decode the motor intent from neural signals obtained from different parts of the nervous system. However, deep neural networks are often computationally complex and not feasible to work in real-time. Here we investigate different approaches' advantages and disadvantages to enhance the deep learning-based motor decoding paradigm's efficiency and inform its future implementation in real-time. Our data are recorded from the amputee's residual peripheral nerves. While the primary analysis is offline, the nerve data is cut using a sliding window to create a "pseudo-online" dataset that resembles the conditions in a real-time paradigm. First, a comprehensive collection of feature extraction techniques is applied to reduce the input data dimensionality, which later helps substantially lower the motor decoder's complexity, making it feasible for translation to a real-time paradigm. Next, we investigate two different strategies for deploying deep learning models: a one-step (1S) approach when big input data are available and a two-step (2S) when input data are limited. This research predicts five individual finger movements and four combinations of the fingers. The 1S approach using a recurrent neural network (RNN) to concurrently predict all fingers' trajectories generally gives better prediction results than all the machine learning algorithms that do the same task. This result reaffirms that deep learning is more advantageous than classic machine learning methods for handling a large dataset. However, when training on a smaller input data set in the 2S approach, which includes a classification stage to identify active fingers before predicting their trajectories, machine learning techniques offer a simpler implementation while ensuring comparably good decoding outcomes to the deep learning ones. In the classification step, either machine learning or deep learning models achieve the accuracy and F1 score of 0.99. Thanks to the classification step, in the regression step, both types of models result in a comparable mean squared error (MSE) and variance accounted for (VAF) scores as those of the 1S approach. Our study outlines the trade-offs to inform the future implementation of real-time, low-latency, and high accuracy deep learning-based motor decoder for clinical applications.
Introduction
For individuals with limb loss, the ultimate goal of neuroprosthetics is to restore natural, intuitive control of a robotic hand. Achieving this requires a system that is both dexterous—capable of moving individual fingers—and real-time, with minimal latency between intent and action. While deep learning (DL) has emerged as a powerful tool for decoding neural signals, its high computational requirements often make it impractical for portable, battery-powered prosthetic devices. This research, published in Frontiers in Neuroscience, addresses this "computational tax" by optimizing DL architectures for peripheral nerve signal decoding.
Experimental Methodology and Data Acquisition
The study focused on decoding motor intent from the peripheral nerves of a single transradial (below-elbow) amputee. Data was recorded via 16 channels using four FAST-LIFE microelectrode arrays implanted in the median and ulnar nerves. To establish ground truth for intended movements, researchers employed a "mirrored bilateral training" paradigm. In this setup, the participant moved their intact hand while wearing a data glove, simultaneously imagining the same movements in their missing limb.
The resulting dataset consisted of over 50 minutes of neural recordings. To simulate real-time conditions, the data was processed into a "pseudo-online" format using sliding 4-second windows.
Decoding Strategies: One-Step vs. Two-Step
The researchers compared two primary architectural approaches:
- One-Step (1S) Strategy: A single model, such as a Recurrent Neural Network (RNN) or Convolutional Neural Network (CNN), predicts the trajectories of all 10 degrees of freedom (DOF) for the fingers simultaneously.
- Two-Step (2S) Strategy: This approach utilizes a "gatekeeper" classification model to first identify which fingers are active. Subsequently, separate regression models are activated only for those specific fingers to predict their trajectories.
A critical innovation in this study was the move away from raw neural data. Instead, the models were fed 14 hand-picked temporal features (e.g., Waveform Length, Zero Crossing). This targeted feature extraction served as a massive dimensionality reduction step, allowing for "shallower" models that are significantly faster and more energy-efficient.
Key Insights and Results
The results highlight the efficacy of intelligent feature engineering and modular architectures. By using 14 specific features, the researchers reduced the number of learnable parameters from approximately 25.9 million in previous state-of-the-art models to just 1.5 million—a 94% reduction in complexity without sacrificing performance.
In terms of accuracy, the classification step in the 2S approach achieved near-perfect results, with both CNNs and Random Forests reaching an F1 score of 0.99. For trajectory prediction (regression), RNNs proved superior due to their ability to handle time-series data, providing smooth and fluid finger movements. The 1S RNN approach achieved high Variance Accounted For (VAF) scores, often exceeding 0.9 for certain degrees of freedom.
Discussion and Future Directions
The study demonstrates that shallow deep learning models, when combined with smart pre-processing, can handle complex motor intent decoding with low latency. The 2S strategy is particularly promising for hardware implementation, as it allows the system to deactivate regression models for inactive fingers, thereby conserving battery life.
However, challenges remain. The current findings are based on a single participant, and neural signals are known to vary significantly between individuals and over time due to physiological changes or electrode drift. Future research must focus on long-term stability and true closed-loop real-time testing, where the decoder is used to control a physical robotic hand in daily living tasks. This work provides a vital roadmap for translating complex "black box" models into efficient, clinical-grade neuroprosthetic controllers.
Summary updated