Technical Architecture

NeuroFeel Technical Frameworks

Explore the technical architecture and methodologies behind our emotion recognition frameworks

WESAD Personalization Framework Architecture

Technical implementation of our adaptive model selection system

Model Architecture

The framework implements a four-tiered model architecture with the following components:

Base Model

Neural network trained on data from all subjects except the target subject using leave-one-out validation

Personal Model

Transfer learning model initialized with base model weights and fine-tuned with minimal subject data

Ensemble Model

Weighted combination of base and personal models with optimal weights learned from validation data

Adaptive Model

Dynamic selection between base and personal models based on confidence thresholds

Technical Innovations

Diversity-aware Calibration

Uses k-means clustering to select the most representative and diverse calibration examples for each emotion class, maximizing information content with minimal data

Confidence-based Selection

Novel algorithm that analyzes prediction confidence from both models to dynamically switch between them during inference

threshold = 0.65 for optimal performance based on validation studies

Subject-specific Feature Importance

Identifies distinctive physiological patterns for each subject using mutual information criterion

Implementation Details

Base Models

MLPClassifier(hidden_layers=(64, 32), alpha=0.01), with options for RandomForest and SVM

Features

Top 20 statistical, temporal, and energy features selected via mutual information

Segmentation

8400 samples per window with 4200 sample overlap (50%)

Calibration

5 examples per emotion class, selected via diversity-aware clustering

Validation

Temporal train/test split (70%/30%) to preserve temporal patterns

Explore WESAD Framework Details

Technical Results

Performance metrics and analysis

Model Performance Comparison

Base Model86.36%

Personal Model84.51%

Ensemble Model87.52%

Adaptive Model88.16%

F1-Score Comparison

Base Model85.75%

Personal Model84.58%

Ensemble Model87.05%

Adaptive Model87.85%

Most Important Features

chest_emg_iqr0.3365

chest_emg_max0.3248

chest_ecg_min0.3206

chest_emg_std0.3011

chest_emg_energy0.2931

Technical Comparison

Characteristic	WESAD Framework	Cross-Dataset Framework
Primary Challenge	Personalization with minimal data	Domain gap between datasets
Core Architecture	Four-tiered model selection	Ensemble domain adaptation
Key Algorithm	Confidence-based selection	Multi-method adaptation
Target Variables	4 emotion classes	Arousal & valence dimensions
Implementation	EnhancedPersonalizationFramework	CrossDatasetFramework
Best Performance	88.16% accuracy (adaptive)	80.27% accuracy (valence)
Feature Selection	Top 20 via mutual information	9 common mapped features
Base Classifier	Neural Network (MLP)	Ensemble (RF, GB, SVM)

Try Interactive Demo