Embedded Files
Along with my senior design team, I am co-developing a non-invasive wearable device designed to provide real-time "pre-ictal" alerts for individuals with epilepsy. This ongoing project bridges the gap between clinical EEG monitoring and wearable consumer health-tech by integrating custom hardware with a machine learning pipeline. My work has focused heavily on the hardware-to-cloud interface, including the fabrication and testing of custom EEG acquisition PCBs. We have successfully validated our hardware’s ability to capture neural signals through statistical analysis of alpha-wave variance and are currently refining the signal-to-noise ratio to ensure clinical-grade data integrity.
In tandem with the hardware development, we are optimizing a Convolutional Neural Network (CNN)/Multi-layer Perceptron (MLP) to classify seizure states. I have helped establish the wireless communication pathway that transmits data from the microcontroller to a cloud-based server for real-time inference. On the mechanical side, we are currently iterating on ergonomic, 3D-printed enclosures with a focus on a behind-the-ear form factor to target the temporal lobe while maintaining user comfort. While we have established a functional "proof-of-concept" data pipeline, our current efforts are focused on improving the ML’s classification accuracy and reducing communication latency to meet our performance targets for a fully functional, real-time seizure prediction prototype.
Device Block Diagram
The system continuously acquires EEG and PPG signals, which are conditioned through pre-amplification, bandpass filtering, and non-inverting amplification. These signals are fed into a CNN+MLP pipeline that classifies the patient’s status as interictal, pre-ictal, or ictal. If a pre-ictal state is detected, the device wirelessly transmits a "seizure likely" alert to a mobile device via Bluetooth, enabling proactive intervention.
Preliminary Wearable Design
I applied human-centered design and rapid prototyping skills to develop this wearable seizure-monitoring headset. Using SolidWorks, I iterated through three electrode configurations, frontal, occipital, and temporal, to balance signal quality with user comfort. This design features compact, 3D-printed enclosures for dry electrodes and integrated haptic feedback for tactile alerts. My current work focuses on flexible TPU printing to ensure skin-conforming contact behind the ear.
EEG-Acquisition Front End
The device acquires EEG through a multi-stage conditioning pipeline, including pre-amplification/instrumentation, bandpass filtering, and non-inverting amplification. Recent refinements focused on distributing total gain throughout the circuit and adjusting band-cutoff frequencies to improve signal quality and address movement artifact challenges. We are currently comparing our circuit to the EEG-Click from Mikrobus as a "gold-standard" comparison.
Circuit Evaluation
To evaluate whether we were successfully recording EEG activity, we conducted an eyes-open/eyes-closed test using our acquisition system with electrodes placed over the occipital region. When the eyes are closed, power in the alpha band (8–12 Hz) typically increases relative to the eyes-open condition. Therefore, detecting this expected change would provide confidence in our system’s ability to measure meaningful neurological activity.
In the figure below (left panel), a clear alternating pattern in alpha-band power corresponds to periods of eye opening and closing. Additionally, we found a statistically significant difference in alpha power between the two conditions using a paired t-test with Mann–Whitney post hoc testing.
ML Pipeline
While most of my project experience has focused on analog circuitry, microcontroller firmware, mechanical design, and neuroscience research, I have included details of the machine learning pipeline here to provide a complete picture of the system. Our model was trained on a single EEG channel from the CHB-MIT open-source dataset, which includes clearly annotated seizure events to facilitate supervised training. Using a leave-one-out cross-validation approach, we generated the confusion matrix shown to the left. The model currently produces very few false positives, which is encouraging; however, it still frequently confuses the pre-ictal and interictal states. Nevertheless, our prototype successfully collects live EEG data and transmits it wirelessly to a machine learning algorithm that provides predictive insights with a measurable degree of accuracy. This foundation is promising, and we plan to continue refining and improving the system throughout the remainder of the semester.
Moving Forward
We are currently at the midpoint of the semester and have successfully demonstrated the ability to record EEG signals, perform analog-to-digital conversion, and analyze the data wirelessly using a server-based machine learning pipeline.
As the semester progresses, we plan to further train our ML model using open-source datasets to improve pre-ictal prediction accuracy. We also aim to enhance the comfort and signal fidelity of the wearable enclosure and reduce movement artifacts within the acquisition circuitry.
Our midterm presentation, attached on the right, provides additional details on the problem motivation, design requirements, and technical specifications.
Page updated
Google Sites
Report abuse