Neuroxis
Deep Learning Platform for Neural Signal Classification & BCI Research
01 / Project Overview
An end-to-end research toolkit for brain-computer interface development. Neuroxis ingests raw EEG and EMG sensor streams, applies artifact removal and frequency-domain feature extraction, and feeds processed signals into custom CNN and LSTM architectures for real-time cognitive state classification. The platform includes an interactive visualization layer for inspecting waveforms, activation maps, and model confidence scores.
02 / The Challenge & Problem
Real-World Problem Statement
Brain-computer interface research is bottlenecked by fragmented toolchains — researchers use separate tools for signal acquisition, preprocessing, model training, and visualization, creating reproducibility issues and slow iteration cycles. Existing frameworks lack integrated real-time inference for closed-loop BCI prototyping.
03 / The Engineering Solution
Implementation & Architectural Approach
Developed an integrated Python pipeline using MNE-Python for EEG preprocessing, custom TensorFlow models for multi-class cognitive state classification, and a Streamlit-based dashboard for interactive signal inspection. The architecture supports both offline analysis and online streaming modes for real-time BCI applications.
04 / Technical Architecture Flow
MNE-Python EEG Pipeline
Reads raw EEG/EMG streams, applies ICA-based artifact rejection, bandpass filters (delta through gamma), and extracts spectral power features per epoch.
CNN-LSTM Hybrid Classifier
Temporal CNN extracts local spatial patterns from electrode grids; LSTM captures sequential dependencies across epochs for cognitive state prediction.
Streamlit Research Dashboard
Real-time topographic maps, waveform browsers, confusion matrices, and Grad-CAM activation overlays for model interpretability.
05 / Key Project Features
Artifact-Free Preprocessing
ICA decomposition automatically identifies and removes eye-blink and muscle artifacts from raw EEG recordings with configurable thresholds.
Multi-Class Cognitive Classification
Classifies mental states (focus, relaxation, motor imagery, fatigue) with 87% accuracy across a 4-class paradigm.
Real-time Streaming Mode
Processes live LSL (Lab Streaming Layer) data streams for closed-loop feedback applications at 250Hz sample rates.
06 / Engineering Challenges & Mitigations
EEG signals are highly non-stationary, causing trained models to degrade rapidly when used across different recording sessions.
Implemented domain adaptation using alignment techniques that normalize session-specific covariance shifts before inference, improving cross-session accuracy by 23%.
Real-time inference latency exceeded the 50ms threshold required for responsive BCI feedback.
Quantized the LSTM layers to INT8 precision with TensorFlow Lite, achieving 12ms inference latency on CPU without accuracy loss.
07 / Technical & Personal Learnings
Developed deep expertise in computational neuroscience, EEG signal characteristics, and frequency-domain feature engineering for time-series classification.
Mastered TensorFlow model optimization techniques including quantization, pruning, and TFLite deployment for edge inference.