The Impact of Training Data Shortfalls on Safety of AI-based Clinical Decision Support Systems

Decision support systems with Artificial intelligence (AI) and specifically Machine Learning (ML) components present many challenges when assuring trust in operational performance, particularly in a safety-critical domain such as healthcare. During operation the Human in/on The Loop (HTL) may need assistance in determining when to trust the ML output and when to override it, particularly to prevent hazardous situations. In this paper, we consider how issues with training data shortfalls can cause varying safety performance in ML. We present a case study using an ML-based clinical decision support system for Type-2 diabetes related co-morbidity prediction (DCP). The DCP ML component is trained using real patient data, but the data was taken from a very large live database gathered over many years, and the records vary in distribution and completeness. Research developing similar clinical predictor systems describe different methods to compensate for training data shortfalls, but concentrate only on fixing the data to maximise the ML performance without considering a system safety perspective. This means the impact of the ML's varying performance is not fully understood at the system level. Further, methods such as data imputation can introduce a further risk of bias which is not addressed. This paper combines the use of ML data shortfall compensation measures with exploratory safety analysis to ensure all means of reducing risk are considered. We demonstrate that together these provide a richer picture allowing more effective identification and mitigation of risks from training data shortfalls.