Privacy Preserving Machine Learning in Healthcare

Abstract

The population of older individuals requiring care in Danish society is increasing.
This trend necessitates innovative and more efficient methods of functioning within the healthcare industry. Fall prevention is an issue requiring attention. To avoid falls, it is essential to evaluate the risk of falling among senior individuals. Should the danger be elevated, the implementation of additional assistance equipment or training programs may commence. Currently, the fall risk assessment is a manual and ineffective process. The optimization of this approach via the use of machine learning techniques delineates the scope of this thesis.

Machine learning employs extensive datasets throughout the training process. Under typical circumstances, this is not an issue; nevertheless, within the context of this thesis, the training data is confidential and so has sensitive characteristics. This complicates the optimization of machine learning and is the fundamental issue addressed in this thesis. What methods may be used to train machine learning algorithms on sensitive datasets?

This thesis assesses the use of federated learning for training machine learning algorithms on decentralized datasets. Additionally, it examines the use of encryption methods and differential privacy to safeguard machine learning models against the disclosure of sensitive information. Finally, it examines how intricate machine learning models, particularly deep learning models, might be rendered explainable to facilitate clearer communication of the findings to senior citizens.   
Numerous tests have shown that federated learning provides a technique for decentralized learning, although at the expense of model performance. Both encryption approaches and differential privacy may enhance the security of machine learning models against data leakage, but at the expense of model performance and complexity. Finally, it is shown how a technique known as SHAP may assist in calculating SHAP feature values, hence making machine learning models more comprehensible.