Abstract
Closed-loop neural interfaces with on-chip machine learning can detect and suppress disease symptoms in neurological disorders or restore lost functions in paralyzed patients. While high-density neural recording can provide rich neural activity information for accurate disease-state detection, existing systems have low channel counts and poor scalability, which could limit their therapeutic efficacy. This work presents a highly scalable and versatile closed-loop neural interface SoC that can overcome these limitations. A 256-channel time-division multiplexed (TDM) front end with a two-step fast-settling mixed-signal DC servo loop (DSL) is proposed to record high-spatial-resolution neural activity and perform channel-selective brain-state inference. A tree-structured neural network (NeuralTree) classification processor extracts a rich set of neural biomarkers in a patient-and disease-specific manner. Trained with an energy-aware learning algorithm, the NeuralTree classifier detects the symptoms of underlying disorders (e.g., epilepsy and movement disorders) at an optimal energy-accuracy tradeoff. A 16-channel high-voltage (HV) compliant neurostimulator closes the therapeutic loop by delivering charge-balanced biphasic current pulses to the brain. The proposed SoC was fabricated in 65-nm CMOS and achieved a 0.227-μJ/class energy efficiency in a compact area of 0.014 mm²/channel.
Introduction: Overcoming the Scalability Wall
Neurological disorders such as epilepsy and Parkinson’s disease affect millions of people worldwide, many of whom do not respond to traditional pharmacological treatments. Closed-loop neuromodulation—a technique where a device monitors brain activity and delivers targeted electrical stimulation to correct pathological states—offers a promising alternative. However, current systems face a significant "scalability wall." High-fidelity detection of brain states requires high spatial resolution (many recording channels), but increasing channel counts typically leads to prohibitive increases in power consumption and silicon area. Most existing systems-on-chip (SoCs) are limited to 8–32 channels and are often specialized for a single task, such as seizure detection.
System Architecture and Technical Innovations
The NeuralTree SoC addresses these challenges through a combination of hardware-algorithm co-design and novel circuit techniques. The system integrates 256 recording channels and 16 stimulation channels into a compact 3.48 mm² active area.
Key innovations include:
- Time-Division Multiplexed (TDM) Front-End: To minimize area, the design shares a single high-performance Analog Front-End (AFE) across 64 channels. This approach significantly reduces the footprint compared to traditional per-channel amplifier designs.
- Two-Step DC Servo Loop (DSL): A major hurdle in TDM is Electrode DC Offset (EDO). Switching between electrodes can cause voltage jumps that saturate amplifiers. The researchers developed a two-step mixed-signal loop that provides "coarse" cancellation in just 7.8 μs followed by "fine" tracking, enabling seamless high-speed channel switching.
- The NeuralTree Classifier: At the core of the SoC is a tree-structured neural network. Unlike standard decision trees that use axis-aligned splits, NeuralTree utilizes "oblique" splits—essentially mini-neural networks at each node. This allows a single tree to achieve the classification power of a much larger random forest.
- Energy-Aware Learning: The classifier is trained using an algorithm that penalizes the use of computationally "expensive" features. This forces the system to prioritize high-accuracy, low-power biomarkers, optimizing the energy-accuracy tradeoff.
- High-Voltage Stimulator: A 16-channel stimulator closes the loop, using a stacked architecture to safely deliver charge-balanced pulses within a standard 65-nm CMOS process.
Experimental Results and Validation
The SoC was extensively validated using human electroencephalography (EEG), intracranial EEG (iEEG), and local field potentials (LFPs), as well as in-vivo rat models.
- Epilepsy Detection: The system achieved 95.6% sensitivity and 96.8% specificity on human EEG datasets.
- Parkinson’s Disease: For the first time in literature, on-chip classification of rest-state tremors from human LFPs was demonstrated with 82.6% sensitivity.
- Motion Decoding: The chip successfully decoded finger movements from human ECoG data with 73.3% accuracy across six classes.
- Efficiency: The system achieved a record-breaking energy efficiency of 0.227 μJ per classification.
Conclusion
NeuralTree represents a significant leap in neural interface technology, providing the scalability and versatility required for next-generation neuroprosthetics and disease management. By identifying and processing only the most informative channels, the system maintains high accuracy while minimizing power. Future work will focus on expanding "online" in-vivo testing and further increasing channel density to support high-fidelity brain-machine interfaces.
Summary updated