Data-driven resiliency assessment of medical cyber-physical systems

Advances in computing, networking, and sensing technologies have resulted in the ubiquitous deployment of medical cyber-physical systems in various clinical and personalized settings. The increasing complexity and connectivity of such systems, the tight coupling between their cyber and physical components, and the inevitable involvement of human operators in supervision and control have introduced major challenges in ensuring system reliability, safety, and security. This dissertation takes a data-driven approach to resiliency assessment of medical cyber-physical systems. Driven by large-scale studies of real safety incidents involving medical devices, we develop techniques and tools for (i) deeper understanding of incident causes and measurement of their impacts, (ii) validation of system safety mechanisms in the presence of realistic hazard scenarios, and (iii) preemptive real-time detection of safety hazards to mitigate adverse impacts on patients. We present a framework for automated analysis of structured and unstructured data from public FDA databases on medical device recalls and adverse events. This framework allows characterization of the safety issues originated from computer failures in terms of fault classes, failure modes, and recovery actions. We develop an approach for constructing ontology models that enable automated extraction of safety-related features from unstructured text. The proposed ontology model is defined based on device-specific human-in-the-loop control structures in order to facilitate the systems-theoretic causality analysis of adverse events. Our large-scale analysis of FDA data shows that medical devices are often recalled because of failure to identify all potential safety hazards, use of safety mechanisms that have not been rigorously validated, and limited capability in real-time detection and automated mitigation of hazards. To address those problems, we develop a safety hazard injection framework for experimental validation of safety mechanisms in the presence of accidental failures and malicious attacks. To reduce the test space for safety validation, this framework uses systems-theoretic accident causality models in order to identify the critical locations within the system to target software fault injection. For mitigation of safety hazards at run time, we present a model-based analysis framework that estimates the consequences of control commands sent from the software to the physical system through real-time computation of the system’s dynamics, and preemptively detects if a command is unsafe before its adverse consequences manifest in the physical system. The proposed techniques are evaluated on a real-world cyber-physical system for robot-assisted minimally invasive surgery and are shown to be more effective than existing methods in identifying system vulnerabilities and deficiencies in safety mechanisms as well as in preemptive detection of safety hazards caused by malicious attacks