A hiring-focused showcase of an EEG affective-computing project: from noisy physiological signals to model comparison and interpretable results, with the larger question made explicit: why emotion-aware AI matters and where it could go next.
71-74%
neural network accuracy range
664
PCA components after reduction
4
emotion dimensions evaluated
The why
A lot of AI work optimizes the machine side of interaction: prediction, speed, automation, and output quality. This project looks at the harder human side: can a model detect useful emotional-state signals from messy physiology, and can we be honest about what the evidence does and does not support?
The practical value is not claiming to read minds.
The value is building AI systems that can reason about uncertainty, adapt to human context, and avoid treating people like identical input devices.
Most software treats emotional state as something users must type, click, or explain. EEG-based affective computing asks whether physiological signals can help systems notice workload, engagement, or stress without relying only on self-report.
EEG is weak, noisy, person-specific, and easy to overclaim. That makes it a useful test bed for practical AI judgement: preprocessing, baselines, evaluation design, and honest communication matter as much as model choice.
If handled responsibly, emotion-aware models can support adaptive learning, assistive interfaces, human-robot interaction, and wellbeing research. The point is not mind reading; it is better human-centred feedback loops.
What it can lead to
The path from notebook to product needs validation, consent, and calibration. These are plausible directions, not claims of deployment readiness.
Tools that adjust pacing, difficulty, or intervention timing when signals suggest overload or disengagement.
Robots or agents that respond to affective context instead of treating every interaction as emotionally neutral.
Dashboards for studying attention, arousal, or recovery patterns, with consent and clear limits around interpretation.
A way to discuss ethics, privacy, calibration, and uncertainty before deploying models around sensitive human signals.
For AI hiring teams
The useful signal is not just the final score. It is the ability to frame an ambiguous data problem, choose defensible baselines, and explain where model performance is limited by the data.
EEG is noisy, high-dimensional, and variable across people. The project shows practical preprocessing and evaluation discipline.
The work does not just present a neural network. It benchmarks simpler models and explains why they underperform.
The report connects model results to constraints hiring teams care about: data quality, scaling, interpretability, and generalisation.
ML workflow
The project moved through dataset selection, EEG preprocessing, feature extraction, dimensionality reduction, and comparative supervised learning.
Feature engineering
PCA retained 90% of variance while reducing the feature set to 664 components for more practical modelling.
DEAP
32 EEG channels, 21.3 hours of data, plus peripheral signals and emotion ratings.
DREAMER
14 EEG channels, consumer-grade headset data, and a separate notebook prototype.
DREAMER notebook
2,132
Rows
2,548
Features
3
Classes
2.15M
GRU params
The notebook used a Keras GRU prototype, 70/30 split, sparse categorical cross-entropy, and early stopping.
Dataset examples
The raw dataset is not a simple table of feelings. Each example is a structured measurement: a stimulus window, many EEG channels, extracted signal features, and emotion labels created from human ratings.
Main report dataset
A participant watches an affective video stimulus while EEG and peripheral signals are recorded. The trial is then paired with self-reported emotion ratings that become the supervised learning target.
Channels
32 EEG channels plus peripheral signals
Window
Video stimulus window, then rating
Label
Valence, arousal, dominance, liking
Example EEG signal sketch
normalized
Channel map
32 nodes
Label space
Feature view
32
EEG channels
21.3h
Recorded data
4
Targets
664
PCA features
pleasant
Valence
activated
Arousal
high
Dominance
positive
Liking
Model comparison
Results are from the current report table for DEAP, organized by classifier and emotion dimension.
Neural Network on Valence
71.01%
Accuracy
73.45%
Precision
76.35%
Recall
74.87%
F1
Best overall performer across emotional dimensions, validating nonlinear modelling for high-dimensional EEG features.
Potential future work
The current report is clear about limitations: unimodal EEG has generalisation constraints, Dominance and Liking are less reliable than Valence and Arousal, and richer architectures would be needed for robust temporal modelling. The next stage should turn those limits into explicit tests.
Stronger generalisation evidence
Move beyond random row splits by holding out entire participants. That would test whether the model transfers to someone it has never seen.
More robust affect estimation
Combine EEG with facial expression, peripheral physiology, or interaction context so the system is not forced to infer everything from one noisy modality.
Better use of signal dynamics
Use LSTMs, temporal CNNs, transformers, or self-supervised EEG representations to learn how emotional state changes through the trial.
Safer product path
Expose confidence, calibration, opt-in boundaries, and failure modes before any real-time affective interface is considered.
The bigger research arc
A stronger version would not just chase a higher score. It would ask whether affective signals remain useful across new people, new tasks, new sensors, and real-time constraints, while making uncertainty visible to users.