Methods for analysing care pathways: ontology, representation, and process perspectives on health data

Thesis

Every interaction with the healthcare system leaves some kind of trace in the form of data, which forms a valuable resource for research. Modern healthcare is about more than individual interventions and diagnoses: equally important is the combination, ordering and timing of events, and the way that the patient moves through the system: the patient pathway. This is the source of a great many research questions, which are difficult to answer: patients cannot be neatly divided into groups, and “compliance” with a particular standard is hard to measure when there are many decision points and unseen variables.

I first examine the extent to which contextual factors affect treatment decisions, and demonstrate that pathway data is shaped by human practices, processes and biases. I then consider the logic behind which data is included in process analysis, and propose an approach that uses ontological knowledge to infer relationships between diagnoses and procedures.

I then propose embedding-based dynamic time warping (E-DTW), an algorithm for describing the similarity between two patients’ pathways. This algorithm is designed with practical characteristics in mind: it incorporates information on both the semantic similarity of the events in a pathway and their temporal patterns, it uses knowledge from standard and publicly available ontologies, and its embeddings can be re-used for different tasks.

Finally, I extend the E-DTW method to measure the semantic similarity between a patient’s pathway and a pathway as laid down in guidelines; in the process, I describe a set of steps for assessing the gap between a guideline and given dataset, and a notation for encoding pathway guidelines in computable form.

Structured data is, by virtue of the way it is recorded and encoded, rich in semantics that can be exploited to create useful insights. Pathways are inherent complex, and purely logical or statistical attempts to analyse them have drawbacks. This thesis combines the use of modern and flexible machine learning concepts with grounding in ontological knowledge, describing a set of methods that allow the benefits of health data to be realised whilst also ensuring that analysis is relevant, reliable, and reproducible.

Authors: Dwyer, OP

• DOI: 10.5287/ora-nbdxeej6q
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: health data science; ontologies; health informatics; knowledge graphs; machine learning; electronic health records; patient pathway; colorectal cancer
• Subjects: Medical informatics