Doctor of Philosophy
Chuah, Mooi Choo
Affordable cloud computing technologies allow users to efficiently outsource, store, and manage their Personal Health Records (PHRs) and share with their caregivers or physicians. With this exponential growth of the stored large scale clinical data and the growing need for personalized care, researchers are keen on developing data mining methodologies to learn efficient hidden patterns in such data. While studies have shown that those progresses can significantly improve the performance of various healthcare applications for clinical decision making and personalized medicine, the collected medical datasets are highly ambiguous and noisy. Thus, it is essential to develop a better tool for disease progression and survival rate predictions, where dataset needs to be cleaned before it is used for predictions and useful feature selection techniques need to be employed before prediction models can be constructed. In addition, having predictions without explanations prevent medical personnel and patients from adopting such healthcare deep learning models. Thus, any prediction models must come with some explanations. Finally, despite the efficiency of machine learning systems and their outstanding prediction performance, it is still a risk to reuse pre-trained models since most machine learning modules that are contributed and maintained by third parties lack proper checking to ensure that they are robust to various adversarial attacks. We need to design mechanisms for detection such attacks. In this thesis, we focus on addressing all the above issues: (i) Privacy Preserving Disease Treatment & Complication Prediction System (PDTCPS): A privacy-preserving disease treatment, complication prediction scheme (PDTCPS) is proposed, which allows authorized users to conduct searches for disease diagnosis, personalized treatments, and prediction of potential complications. (ii) Incentivizing High Quality Crowdsourcing Data For Disease Prediction: A new incentive model with individual rationality and platform profitability features is developed to encourage different hospitals to share high quality data so that better prediction models can be constructed. We also explore how data cleaning and feature selection techniques affect the performance of the prediction models. (iii) Explainable Deep Learning Based Medical Diagnostic System: A deep learning based medical diagnosis system (DL-MDS) is present which integrates heterogeneous medical data sources to produce better disease diagnosis with explanations for authorized users who submit their personalized health related queries. (iv) Attacks on RNN based Healthcare Learning Systems and Their Detection & Defense Mechanisms: Potential attacks on Recurrent Neural Network (RNN) based ML systems are identified and low-cost detection & defense schemes are designed to prevent such adversarial attacks. Finally, we conduct extensive experiments using both synthetic and real-world datasets to validate the feasibility and practicality of our proposed systems.
Xue, Qinghan, "Effective and Secure Healthcare Machine Learning System with Explanations Based on High Quality Crowdsourcing Data" (2019). Theses and Dissertations. 4375.