Embracing Machine Learning in Safety Assurance in Healthcare

Machine learning (ML) is becoming more widely used in many different sectors, including automotive, aviation and healthcare. ML has a great potential to change society and to improve peoples' lives. However, the prospect of ML also poses many challenges; one of the biggest challenges is safety. Thus, there are two important questions that require urgent answers: (1) Are well-established safety engineering methods still appropriate and effective in assuring the safety of ML in some representative healthcare scenarios? (2) Are there new opportunities for well-established safety engineering methods with the development of ML and why are they specifically good for safety in this domain?

In this thesis, the first question is explored from the viewpoint of designing ML models. The second question is explored from two perspectives: explanaibility of ML models in support of safety assurance; and using ML to update safety analysis. Both these questions are addressed in the context of healthcare. In other words, this thesis investigates how ML can be embraced in the safety assurance of healthcare applications.

Through exploration of three concrete clinical case studies, the thesis demonstrates that well-established safety engineering methods can be applied to ML systems to integrate safety into their design process in healthcare. It further identifies different ways in which ML can assist well-established safety engineering methods, and concludes that there are many opportunities for greater synergy between ML and safety engineering in healthcare and, potentially, in other domains.