Improving Robustness Falsification for Medical Device Software

Abstract

The artificial Pancreas(AP) is a closed-loop system based on the combination of a continuous glucose monitor, a computer-controlled algorithm, and an insulin pump (Blauw et al., 2016). Insulin pumps are designed for continuous delivery of insulin and mimic the way of human pancreas works by delivering small doses of insulin continuously (the basal dose). The device is also used to deliver variable amounts of insulin when a meal is taken (bolus dose). However, the errors in insulin pump software can bring risks to the patients including high blood glucose levels and dangerously low glucose levels (Zhang et al., 2011). In this research, we use Hovorka artificial pancreas mathematical model, which is designed a basic insulin-glucose regulatory system (Artificial Pancreas) in a diabetic patient (Cobelli et al., 1982 Dalla Man et al., 2007 Patek et al., 2009). Patients with glucose over 9 for a long time can cause damage to human health including the eye, kidney, heart, and blood vessels, but glucose under 4.5 may lead to trembling or weak, under 2.5 can lead to coma or even death. We see that glucose under 4.5 is more dangerous than over 9. The research problem is how can we make the software more sensitive to more dangerous risks? In this dissertation, we will use S-TaLiRo (Annpureddy et al., 2011) robustness tool which utilizes optimization-based falsification a method of search-based testing that adopts stochastic optimization. With this falsification approach, the robustness falsification problem is turned into an optimization problem (Fainekos and Pappas, 2006a) (see figure 1.1). Since the lower glucose is dangerous than higher glucose, we are trying to enhance the robustness when the glucose is under 4.5. Let s say the scaling would be used to increase the robustness measure for the glucose is under 4.5, and when the glucose is 4, the original robustness is 4-4.5 = -0.5, after twice enhancing the original robustness, scaled robustness value should be -1.0. We explored 4 different approaches to scale robustness, including scaling during optimize search, simple scaling, discontinuous scaling, and continuous scaling.