Special Edition Blog Series on PhD Abstracts (Part V)

This post is a continuation of the blog post series on PhD abstracts. You can find the first part of the series here.

This research will propose a new approach to protect against risks related to personal data exploitation, drawing a methodology for the implementation of data management and analytics in edge computing and serverless offerings in considering privacy properties to modulate the prevention of risks and promotion of innovation. In addition, it will establish AI-driven processes to increase the user’s ability to define in a more accurate way both his offerings in edge computing environments and the data management and analytics as regards the protection of her/his privacy; draw the architecture in terms of data governance and analytics linked with the resource resources management on such dynamic environments.

 

Christos Magkos: Personal Health Information Management
Systems for User Empowerment

In the era of immense data accumulation in the healthcare sector, effective data management is becoming increasingly relevant in two domains: data empowerment and personalisation. As healthcare shifts towards personalized and precision medicine, prognostic tools that stem from robust modeling of healthcare data, while remaining compliant with privacy regulations and the four pillars of medical ethics (Autonomy, Beneficence, Non-maleficence and Justice) are lacking. The following thesis assesses the principles that the design of health data storage and processing should adhere to through the prism of  personal information management systems (PIMS). PIMS enable decentralized data processing, while adhering to data minimization and allowing for control of data exposure to third parties, hence enhancing privacy and patient autonomy. We propose a system where data is processed in a decentralized fashion, providing actionable recommendations to the user through risk stratification and causal inference modeling of health data sourced from electronic health records and IoT devices. Through an interoperable personal information management system, previously fragmented data which can be variably sourced and present with inconsistencies can be integrated into one system consistent with the EHDS, and as such data processing can proceed more accurately.  When attempting to design clinically actionable healthcare analytics and prognostic tools, one of the main issues arising through current risk stratification models is the lack of actionable recommendations that are deeply rooted to pathologies analyzed. We therefore compare whether causal inference models derived from existing literature and known causal pathways can provide equally accurate predictions to risk stratification models when medical outcomes are known. This would allow for explainable and actionable outcomes, as physicians are reluctant to act upon “black box” recommendations due to medical liabilities and patients are less likely to be compliant to unexplained recommendations, rendering them less effective when translated to the clinic.  Simulated datasets based on different types of data collected are analyzed according to risk stratification and causal inference models in order to infer potential recommendations. Different methodologies of risk stratification and causal inference are assessed and compared in order to find the optimal model that will function as a source of recommendations. Finally, we propose a holistic model under which the user is fully empowered to share data, analytics and metadata derived from this data management system with doctors, hospitals and researchers respectively with recommendations that are designed to be explainable and actionable.