[GT] NeuralTree: A 256-Channel 0.227-¥ìJ/Class Versatile Neural Activity Classification and Closed-Loop Neuromodulation SoC
By Uisub Shin, IEEE JOURNAL OF SOLID-STATE CIRCUITS, November 2022, Vol. 57, Iss.11
Tens of millions of people suffer from Parkinson¡¯s disease and epilepsy. The resulting human and financial cost of is staggering.
Fortunately, researchers have now been able to combine low-power chip design, machine learning algorithms, and soft implantable electrodes to produce NeuralTree, a closed-loop neuromodulation system-on-a-chip that can detect and alleviate symptoms of Parkinson¡¯s disease and epilepsy.
NeuralTree benefits from the accuracy of a neural network and the hardware efficiency of a decision tree algorithm. It¡¯s the first time we¡¯ve been able to integrate such a complex, yet energy-efficient neural interface for seizure or tremor detection, as well as for multiclass tasks such as finger movement classification for neuro-prosthetic applications.
The results of this research were presented at the 2022 IEEE International Solid-State Circuits Conference and published in the IEEE Journal of Solid-State Circuits.
NeuralTree functions by extracting neural biomarkers from brain waves. It then classifies the signals and indicates whether they herald an impending epileptic seizure or Parkinsonian tremor. If a symptom is detected, a neurostimulator - also located on the chip - is activated, sending an electrical pulse to block it. This is the first demonstration of Parkinsonian tremor detection with an on-chip classifier.
The researchers explain that NeuralTree¡¯s unique design gives the system an unprecedented degree of efficiency and versatility compared to the state-of-the-art.
The chip boasts 256 input channels, compared to 32 for previous machine-learning-embedded devices; this allows more high-resolution data to be processed on the implant.
The chip¡¯s extremely small size gives it great potential for scalability to more channels. And the integration of an ¡®energy-aware¡¯ learning algorithm makes NeuralTree highly energy efficient.
The chip¡¯s machine learning algorithm was trained on datasets from both epilepsy and Parkinson¡¯s disease patients, and accurately classified pre-recorded neural signals from both categories.
As a next step, the team is interested in enabling on-chip algorithmic updates to keep up with the evolution of neural signals. That¡¯s important because neural signals change; so, over time the performance of a neural interface will decline unless updated. One way to address that is to enable on-chip updates, or algorithms that can update themselves.