Reproducibility in A.I. for biomedical research: an application to Parkinson's Disease

Deep learning-based data analysis techniques have found many uses in biomedical re- search. Recent expansion of open source databases and computational tools has fostered distributed and explorative research. Under these conditions, reproducibil- ity and experimental rigor must be ensured. Developing explicit analysis pipelines exposes the scientific process and yields reproducible results. In this thesis, we look at the case of deep learning-based data analysis for Parkinsons disease (PD) research. We develop end-to-end pipelines in two PD-related fields: accelerometer data analysis and neuroimage analysis. First, we construct a simple yet robust recurrent neural network for classifying motor activity from accelerometer data alone; this has applications for identifying the mo- tor symptoms of PD. Next, we propose a novel graph convolutional network architecture for distinguishing PD patients vs. healthy controls from multimodal neuroimage data. Our pipelines standardize the data preprocessing and analysis steps, fostering reproducibility and deliberate progression of their respective fields.