Explainable Edge AI EEG Stroke Screening on Raspberry Pico

point-of-care stroke screening using single-chip EEG processing, TinyML on Raspberry Pi Pico, and mobile emergency triage guidance

Comment

Inspiration

Ischemic and hemorrhagic strokes remain primary causes of long-term neurological disability worldwide. Clinical outcomes depend heavily on immediate intervention within the critical time. However, centralized hospital imaging tools like CT and MRI scans are expensive, time-consuming, and unavailable in ambulances or low-resource rural health centers. Furthermore, transmitting continuous, high-bandwidth bio-signals to cloud servers introduces latency, requires stable internet connectivity, and raises patient bio-data privacy concerns. To overcome these barriers, we developed a low-cost, offline, portable Edge AI framework for rapid point-of-care stroke screening using scalp Electroencephalography (EEG).

What it does

Our system is a wearable, zero-cloud stroke screening device that executes end-to-end EEG processing and machine learning inference directly on an embedded microcontroller in Python:

  1. Signal Acquisition: Captures scalp EEG biopotentials via a compact BioAmp EXG Pill analog front-end sampled at 250 Hz by a Raspberry Pi
  2. On-Chip Artifact Mitigation: Applies digital IIR bandpass filtering (0.5Hz- 30Hz) and static Independent Component Analysis (ICA) spatial unmixing directly on-device in Python to suppress ocular and cardiac noise.
  3. Feature Engineering: Computes spectral asymmetry features over 2-second sliding temporal epochs, isolating Delta-to-Alpha Ratio (DAR), Brain Symmetry Index (BSI), and Power Spectral Density (PSD) metrics.
  4. Local Machine Learning Inference:Executes a lightweight Support Vector Machine (SVM) classifier in Python on the RP2040 dual ARM Cortex-M0+ processor for real-time Stroke vs. Non-Stroke detection with sub-10 ms latency.
  5. Mobile Emergency Telemetry: Transmits low-overhead alert payloads via BLE/Serial to a companion app upon anomaly detection to guide non-clinical bystanders through immediate FAST symptom validation, auto-route to the nearest stroke center, and generate clinical bio-signal logs.

How we built it

  1. Hardware Assembly: Interfaced the BioAmp EXG Pill analog front-end directly with the Raspberry Pi Pico microcontroller (RP2040 dual ARM Cortex-M0+ @ 133 MHz).
  2. Embedded Python Pipeline: Developed a Python-native signal processing pipeline leveraging embedded math libraries to handle real-time IIR digital filtering, static ICA matrix operations, and spectral feature computations (DAR, BSI, PSD).
  3. Machine Learning Deployment: Trained an SVM classifier in Python using Scikit-learn on EEG asymmetry datasets, then deployed the trained model parameters into the MicroPython/CircuitPython environment on the Pico to execute predictions efficiently on the microcontroller.
  4. Companion Mobile Application: Created a mobile interface to parse low-overhead serial/Bluetooth telemetry, guide bystanders step-by-step through the emergency FAST protocol, and automate emergency routing.

Challenges we ran into

  1. Python Execution on Resource-Constrained Hardware: Running matrix operations for spatial unmixing (ICA) and feature extraction within the strict memory and fixed-point constraints of the RP2040 required light memory buffers and optimized floating-point computations.
  2. Real-Time Artifact Removal at the Edge: Filtering out high-amplitude eye blinks and cardiac signals without delaying real-time inference or relying on cloud processing.
  3. Asymmetry Metric Calibration: Tuning sliding epoch windows (2 seconds) to ensure reliable spectral feature extraction (DAR and BSI) without exhausting the microcontroller's RAM.

Accomplishments that we're proud of

  1. Sub-10ms Inference on RP2040:Achieving ultra-low-latency stroke classification running entirely in Python directly on the standard Raspberry Pi Pico.
  2. Zero Cloud Dependence:Complete data privacy and full functionality in low-resource or remote offline emergency settings.
  3. Clinical Pathophysiology Mapping; Successfully converting clinical indicators of cerebral hypoperfusion (hemispheric power asymmetry) into a portable, low-cost diagnostic tool.

What we learned

  • Techniques for optimizing digital filtering, matrix unmixing, and feature engineering for resource-constrained ARM Cortex-M0+ processors.
  • How to deploy and execute Scikit-learn machine learning parameters efficiently within MicroPython/CircuitPython environments.
  • The critical role of inter-hemispheric asymmetry metrics (DAR, BSI) in non-invasive stroke screening.

What's next for Explainable Edge AI EEG Stroke Screening on Raspberry Pico

  1. Custom PCB Integration: Designing a single custom PCB board integrating the RP2040 chip, BioAmp front-end, and a LiPo power system for a compact, all-in-one wearable form factor.
  2. Expanded Multi-Class Classification: Expanding the ML framework to differentiate between Ischemic vs. Hemorrhagic stroke types and transient ischemic attacks (TIA).
  3. Clinical Validation Trials:Partnering with clinical researchers to validate on-device accuracy against multi-channel clinical EEG datasets in prehospital ambulance environments.

Built With

Share this project:

Updates

Submission history