In this project, we developed deep learning architectures for Parkinson’s disease stratification across high-dimensional multi-omics datasets, achieving over 90% accuracy. The work was carried out during the first year of my master’s as a Machine Learning Research Assistant at Turku BioScience.
A central idea was leveraging the spatial inductive biases of convolutional neural networks on tabular biological data by investigating tabular-to-image transformations such as DeepInsight and IGTD. We also studied how explainability frameworks, e.g., SHAP, could aid in identifying candidate biomarkers subject to biological fidelity.
We presented a poster at the EDISS 2024 Winter School at Mälardalen University, Västerås, Sweden.