[Paper Review] A Portable, Self-Contained Neuroprosthetic Hand with Deep Learning-Based Finger Control

Anh Tuan Nguyen, Markus W. Drealan+7 more|arXiv (Cornell University)|Mar 24, 2021

Muscle activation and electromyography studies38 references45 citations

TL;DR

This paper presents a portable, self-contained neuroprosthetic hand system that uses a deep learning-based recurrent neural network (RNN) decoder deployed on an NVIDIA Jetson Nano edge computing platform to enable real-time, high-accuracy (95–99%) control of individual finger movements from peripheral nerve signals in a transradial amputee. The system achieves low-latency (50–120 ms) performance across diverse environments, demonstrating the feasibility of embedding AI-driven neural decoding in wearable, clinical-grade neuroprostheses.

ABSTRACT

Objective: Deep learning-based neural decoders have emerged as the prominent approach to enable dexterous and intuitive control of neuroprosthetic hands. Yet few studies have materialized the use of deep learning in clinical settings due to its high computational requirements. Methods: Recent advancements of edge computing devices bring the potential to alleviate this problem. Here we present the implementation of a neuroprosthetic hand with embedded deep learning-based control. The neural decoder is designed based on the recurrent neural network (RNN) architecture and deployed on the NVIDIA Jetson Nano - a compacted yet powerful edge computing platform for deep learning inference. This enables the implementation of the neuroprosthetic hand as a portable and self-contained unit with real-time control of individual finger movements. Results: The proposed system is evaluated on a transradial amputee using peripheral nerve signals (ENG) with implanted intrafascicular microelectrodes. The experiment results demonstrate the system's capabilities of providing robust, high-accuracy (95-99%) and low-latency (50-120 msec) control of individual finger movements in various laboratory and real-world environments. Conclusion: Modern edge computing platforms enable the effective use of deep learning-based neural decoders for neuroprosthesis control as an autonomous system. Significance: This work helps pioneer the deployment of deep neural networks in clinical applications underlying a new class of wearable biomedical devices with embedded artificial intelligence.

Motivation & Objective

• To address the clinical challenge of deploying computationally intensive deep learning neural decoders in portable, real-time neuroprosthetic systems.
• To overcome the limitations of traditional low-computational methods (e.g., KNN, Kalman filters) in achieving dexterous, intuitive control of prosthetic hands.
• To demonstrate the feasibility of embedding a full deep learning inference pipeline—training, deployment, and real-time decoding—on a compact, edge-based platform suitable for long-term clinical use.
• To integrate neural decoding with a fully self-contained, wearable system that can be attached to existing prosthetic sockets.
• To enable intuitive, multi-finger control in real-world and dynamic environments through on-device processing.

Proposed method

• The neural decoder is based on a recurrent neural network (RNN) architecture trained to decode motor intent from intraneural recordings of peripheral nerve signals (ENG) acquired via implanted intrafascicular microelectrodes.
• The RNN model is deployed on the NVIDIA Jetson Nano, a compact edge computing device with GPU acceleration, enabling real-time inference without reliance on external servers.
• A custom printed circuit board (PCB) integrates the Jetson Nano with neural recording and stimulation electronics, including Neuronix microchips for bidirectional neural interfacing.
• The system processes 1-second windows of nerve data with overlapping segments to minimize latency and supports real-time control of individual finger movements.
• The entire system is powered by a 7.4V, 2,200mAh Li-ion battery and weighs 210g, allowing it to be worn as a self-contained unit on a transradial prosthetic socket.
• Software stack includes CUDA-enabled deep learning frameworks (TensorFlow, PyTorch) and optimized pre-processing pipelines for filtering and feature extraction.

Experimental results

Research questions

• RQ1 Can a deep learning-based neural decoder be effectively deployed on a portable, edge computing platform for real-time neuroprosthetic control?
• RQ2 Can such a system achieve high accuracy and low latency in both laboratory and real-world environments using peripheral nerve signals from an amputee?
• RQ3 How does the performance of the embedded RNN decoder compare to conventional decoding methods in terms of dexterity and responsiveness?
• RQ4 What are the practical limitations of current edge hardware (e.g., Jetson Nano) in supporting complex, real-time deep learning inference for neuroprosthetics?
• RQ5 To what extent can on-device processing enable long-term, autonomous use of AI-powered neuroprostheses?

Key findings

• The system achieved 95–99% accuracy in decoding individual finger movements from peripheral nerve signals in a transradial amputee during real-time control.
• The system demonstrated low-latency performance of 50–120 milliseconds from nerve signal acquisition to motorized finger movement initiation.
• The system remained robust across various arm and body postures, as well as in both laboratory and real-world environments over several hours of continuous operation.
• The amputee reported that the system felt intuitive and natural, comparing it to the function of a biological hand, and expressed strong confidence in its potential for everyday use.
• The integration of the Jetson Nano enabled full on-device processing, eliminating dependency on external computing and enabling true portability and self-containment.
• The system’s performance was maintained despite hardware constraints, indicating that modern edge platforms can support complex deep learning models for clinical neuroprosthetics.