Interpretability of Gait Features of HSP Patients in Deep Neural Networks

Master’s Thesis

Abstract

The application of machine learning (ML) can help to improve medical diagnoses and treatments. However, in diagnoses and treatments, not only the prediction of a disease is important but the information about relevant features for the prediction. To make “black box” machine learning techniques like deep neural networks (DNN) interpretable is crucial in medical predictions. In this project the Layer-wise relevance proagation (LRP), a backward propagation technique for explanation, will be used to interpret the unique gait patterns in spastic patients. The gait of patients with hereditary spastic paraplegia (HSP) and healthy controls were recorded in a movement laboratory. HSP is a hereditary slowly progressive neurological movement disorder characterized by spastic gait. To predict whether a subject is affected by HSP we want to use DNNs and the LRP to predict and interpret the gait of healthy controls, HSP patients and preclinical subjects who do not show typical clinical symptoms yet.

Task

The aim of this work is to implement and test DNNs and the Layer-wise relevance propagation to predict and interpret the gait differences in HSP patients and preclinical subjects. The relevance of different features should be made accessible and interpretable to improve medical diagnoses and treatments.

Requirements

  • Interest in neuroscience
  • Interest in Machine Learning
  • Experience in Matlab / Python

Contact

Niethammer, Christian