Portiloop: Deep Learning Closed-Loop EEG System

• Portiloop is a portable closed-loop brain stimulation system that uses deep learning to analyze EEG data and trigger precise neural interventions.
• It features a modular pipeline with real-time signal processing, low latency (approximately 300 ms), and cost-effective hardware under $300 USD.
• The open science initiative offers complete open-hardware and open-software resources, fostering reproducible research and community collaboration.

Portiloop is a deep learning-based, portable, and low-cost system for closed-loop brain stimulation designed to target specific brain oscillations using real-time electroencephalography (EEG) signal analysis and event-triggered stimulation. Engineered as an open science initiative, it provides reproducible open-hardware and open-software resources to facilitate research in fundamental and clinical neuroscience applications requiring closed-loop paradigms. The architecture features a complete acquisition-to-stimulation pipeline, real-time neural event detection using an embedded compact artificial neural network (ANN), and supports low-latency auditory or electrical stimulation modalities (Valenchon et al., 2021).

1. System Architecture and Signal Processing Pipeline

Portiloop implements a modular closed-loop pipeline designed for real-time detection and stimulation of neural events. The sequence is as follows:

1. EEG Electrode Acquisition: Scalp electrodes capture neural activity.
2. Analog Front-End and Digitization: Signals are conditioned and digitized with an ADS1299 ADC.
3. Embedded Real-Time Computation: Processing is performed on an embedded platform (Xilinx Pynq Z2 or Google Coral SoM). Core stages include:
   • Finite impulse response (FIR) band-pass filtering (0.5–30 Hz) and 50/60 Hz notch filtering.
   • Feature computation via spindle-band (12–16 Hz) envelope extraction, squaring, and moving average.
   • Exponential moving-average-based signal standardization.
   • Forward pass through a lightweight CNN+RNN neural network for event (e.g., spindle) detection.
4. Event Logic and Stimulation: Event detection triggers a single stimulus per neural event using audio or other outputs directed at the subject.
5. External Interface: Configuration, logging, and real-time LSL streaming via USB or Wi-Fi.

The pipeline is optimized for minimal detection-stimulation latency (approximately 300 ms end-to-end). All resources are openly available at https://github.com/Portiloop (Valenchon et al., 2021).

2. Hardware and Cost Considerations

Portiloop prioritizes accessibility and flexibility through a carefully selected hardware stack:

• Embedded Computing: Compatible with Xilinx Pynq Z2 (FPGA + ARM) and Google Coral System-on-Module. This supports efficient inference of the compact ANN on low-power, hand-held devices.
• Power Consumption: Total current draw is maintained below 1.5 A, supporting battery-powered operation for portability.
• Component Cost: The cumulative cost of the embedded computer, EEG front-end, and battery is maintained below $300 USD.
• Footprint: The device is hand-held and suitable for mobile or bedside deployment.
• Open Hardware: All PCB designs and assembly instructions use commercially available components, fostering reproducibility and community-driven modification (Valenchon et al., 2021).

3. Deep Learning Model and Hyperparameter Optimization

At the core of Portiloop’s real-time detection capability is a lightweight ANN designed for embedded inference:

• Model Architecture: The network integrates convolutional layers (CNN) for spatial-temporal filtering with recurrent layers (RNN) for sequence modeling.
• Optimization: An exploration algorithm is provided to automatically tune model hyperparameters to the target oscillatory event (e.g., sleep spindles), improving versatility across different EEG phenomena.
• Feature Engineering: The system employs spindle-band envelope features derived from 12–16 Hz filtered data, squared and averaged, to emphasize transient oscillatory events.
• Performance: ANN forward pass duration is subsumed in the total 300 ms detection-stimulation loop latency, aligning with the real-time requirements for closed-loop stimulation (Valenchon et al., 2021).

4. Validation Protocols and Benchmark Results

The validation of Portiloop centers on a challenging benchmark: real-time sleep spindle detection. Key details include:

• Data Set: Evaluation utilizes the Massive Online Data Annotation spindle dataset (MODA), allowing comparison with human expert consensus labels.
• Performance Comparison: The Portiloop achieves real-time event detection accuracy comparable to off-line expert annotation performance, specifically for sleep spindle events.
• Stimulus Control: Event logic enforces a policy of at most one stimulus per detected spindle, minimizing confounds in neurophysiological response measurement.
• Latency Characterization: The measured end-to-end delay (signal acquisition, feature extraction, ANN inference, and stimulation) is approximately 300 ms.

A plausible implication is that Portiloop’s accuracy and latency are suitable for investigating time-locked modulation of oscillatory brain events such as spindles or slow waves in both research and translational contexts (Valenchon et al., 2021).

5. Open Science Initiative and Community Engagement

A fundamental design goal is to promote open science and reproducibility:

• Open-Source Repository: All hardware design files, software source code, and pre-trained model weights are released under permissive open-source licenses.
• Community Collaboration: Extensive documentation and tutorials are provided, facilitating adaptation to novel experimental paradigms or oscillatory targets.
• Interface Support: The system supports the Lab Streaming Layer (LSL) protocol, enabling integration with broader neuroscience experimental workflows.
• Extensibility: Researchers are encouraged to modify and extend both the hardware and software, potentially advancing closed-loop stimulation methodologies for fundamental neuroscience and clinical research (Valenchon et al., 2021).

6. Applications and Prospective Impacts

Portiloop is intended for a range of applications in fundamental and clinical neuroscience where temporally precise intervention based on neural signals is essential:

• Basic Research: Enables studies probing causal relationships between neural events (e.g., sleep spindles) and cognitive or behavioral outcomes via time-locked interventions.
• Clinical Potential: The architecture may support translational research targeting restoration of degraded memory function or other pathologies informed by oscillatory biomarkers.
• Methodological Innovation: Its accessibility, low cost, and compact form may democratize closed-loop approaches, accelerating methodological innovation in both laboratory and point-of-care settings.

The open and modular architecture positions Portiloop as a reference platform for reproducible closed-loop neuroscience, encouraging further community-driven development (Valenchon et al., 2021).