Grant: Discovery Grants Program - Individual

Workflows define a set of processes needed to realize a given task, ranging from registering a customer into a system, over determining eligibility for credit cards, to the evidence-based diagnosis and treatment of patients. However, these is often a disconnect between "idealized", normative workflows and real-world processes as carried out by people in practice. For instance, in healthcare, clinical guidelines define step-by-step processes for diagnosis and treatment based on the latest clinical evidence. In practice, however, real-world care processes often deviate from clinical guidelines, for the following reasons: (1) it is often impossible for clinicians to keep abreast of these guidelines, which may span hundreds of pages and are updated every few years or so; and (2) clinicians may overrule guidelines that lack consideration for complex or atypical cases, such as multiple co-occurring illnesses (multimorbidity) or particular patient characteristics, instead relying on their hard-gained experience for these cases. Knowledge-driven Decision Support (CDS) is one tool to overcoming this disconnect between guidelines and real-world practice. CDS tools are able to recommend next steps based on workflows found in clinical guidelines, meaning clinicians don't have to be 100% up to date with the latest guidelines. However, this does not cover cases where clinicians intentionally deviate from guidelines to cope with atypical cases or minority populations. This research program has the following objectives. Firstly, we aim to discover real-world processes from recorded event data, which is often low-quality and incomplete. Next, we plan to systematically compare discovered processes with normative workflows, identifying deviations and their possible underlying reasons. In doing so, we aim to gain insights into real-world practices - for instance, for coping with atypical cases or person/object characteristics - with the ultimate goal of improving normative "idealized" workflows. We will propose methods that rely on Knowledge Graphs to realize these objectives. Importantly, these methods will be explainable, meaning they will offer human-readable explanations for any choices made during process discovery or comparison. This also enables interactivity with a domain expert during process discovery and comparison, who may choose to accept or reject an automated output based on the explanation. The developed methods will be evaluated on the Rick Hansen Spinal Cord Injury Registry dataset, and the Admission-Discharge-Transfer dataset from The Ottawa Hospital.