View Proposal

---

Proposer

Marta Vallejo

Title

Generative AI for Wearable Finger-Tapping Analysis in Parkinson's Disease (with the collaboration of University of York)

Goal

The project aims to develop and evaluate generative and conventional data augmentation techniques, including Variational Autoencoders (VAEs), Conditional VAEs (CVAEs), TimeGAN, SMOTE and ADASYN, to improve the classification of early Parkinson's disease from wearable finger-tapping sensor data, while assessing their robustness, interpretability and clinical relevance.

Description

Early detection of Parkinson's disease is essential for improving patient management, yet identifying subtle motor impairments remains challenging. This project will investigate how generative AI techniques can improve the analysis of wearable finger-tapping sensor data collected from a real clinical cohort of patients assessed using the MDS-UPDRS. Students will compare multiple data augmentation approaches, including Variational Autoencoders (VAEs), Conditional VAEs (CVAEs), TimeGAN, SMOTE and ADASYN, to determine which techniques best improve classification performance and model robustness. The project will also evaluate the realism of the generated synthetic data, investigate model explainability using feature importance methods such as SHAP, and assess performance at both window and patient levels. The expected outcome is a comprehensive comparison of augmentation strategies for wearable healthcare data, with the aim of producing a publication in a biomedical engineering or digital health journal.

Resources

(1) Real clinical finger-tapping dataset with wearable sensor recordings from Parkinson's disease patients. (2) Preprocessing and feature extraction pipeline developed in the previous student project. (3) Baseline MLP, LSTM, VAE and CVAE implementations available in Python/PyTorch.

Background

Wearable sensors provide an objective and non-invasive way to quantify motor impairment in Parkinson's disease and have become an important source of digital biomarkers. However, machine learning models are often limited by the small size and class imbalance of clinical datasets. This project builds on an existing research pipeline to investigate whether modern generative AI and data augmentation techniques can improve the robustness, generalisation and interpretability of models developed for real-world clinical sensor data.

Url

Difficulty Level

Variable

Ethical Approval

Full

Number Of Students

2

Supervisor

Marta Vallejo

Keywords

parkinson's disease, wearable sensors, finger tapping, generative ai, data augmentation, deep learning, digital biomarkers, explainable ai, clinical datasets, machine learning

Degrees

Bachelor of Science in Computer Science
Master of Engineering in Software Engineering
Master of Science in Artificial Intelligence
Master of Science in Computing (2 Years)
Master of Science in Data Science
Master of Science in Robotics
Master of Science in Software Engineering
Bachelor of Science in Computing Science
Bachelor of Engineering in Robotics
Postgraduate Diploma in Artificial Intelligence