Student Projects

A Streamlit-based machine learning pipeline for automated analysis of psychometric data from the TIMED study. Supports multi-model regression with Optuna hyperparameter tuning and SHAP interpretability for cognitive and neuropsychological assessment.

Can playing Fruit Ninja reveal reading difficulties? This thesis uses eye-tracking metrics collected during gameplay to predict Rapid Automatised Naming scores and screen for developmental dyslexia — without requiring participants to read. Analysed 23 oculomotor features from 57 participants using YOLOv8 object detection and ML classification.

Predicting brain age from sleep EEG recordings using a U-Sleep architecture with foundation model principles. Includes automated sleep staging, feature extraction, age regression, MCI classification, and correlational analyses.

Investigating how drivers' visual attention and motor control adapt as they gain route familiarity. A five-stage pipeline analyses eye-tracking, pedal activity, and steering data from 27 participants in a driving simulator, using CNN-based video segmentation and machine learning to reveal the transition from reactive to anticipatory control strategies.

A multimodal text editor combining eye-tracking gaze input with speech recognition for hands-free text editing. Users calibrate eye tracking, then use gaze to position a cursor and voice commands to dictate and edit text.

Comparing wearable and nearable biosensors against clinical polysomnography for detecting sleep motor activity. Cross-correlation analysis of sensor data to evaluate accuracy of consumer-grade sleep tracking devices.

Bipartite community detection applied to Letterboxd social data using the Bilouvain algorithm. Analyses user-film interaction networks to uncover viewing communities and taste clusters through graph-based methods.

A multimodal AI assistant that lets users ask questions about their environment using gesture recognition, speech-to-text, and object detection. Combines MediaPipe hand tracking, OpenAI Whisper, and Facebook DETR to understand and respond to visual scenes.

Graph-based analysis of the Swiss public transport network using Neo4j and community detection algorithms (Louvain, Girvan-Newman, Leiden). Explores network structure, centrality, and clustering patterns in SBB timetable data.

A cooperative platformer game controlled through three input modalities: hand gesture recognition via MediaPipe, voice commands via Vosk speech recognition, and traditional keyboard input. Built for the Multimodal User Interfaces course.

A real-time neurofeedback system using the Muse 2 EEG headband. Streams and processes brain signals via BrainFlow, computes posterior alpha and frontal beta band powers to derive an arousal index, and applies a cybernetic control loop with hysteresis-based state management. Includes a Streamlit dashboard for live visualisation and post-session analysis.