Developing Digital Twins for Predictive Outcome Analysis in Arthroplasty

Abstract

This work introduces a new concept, namely a digital twin, to understand how clinical data relates to the outcomes and as a next step how it can be put to good beneficial use in clinical decision making. Undergoing surgery is a very invasive experience for patients. Both in the sense of uncertainty before the surgery and also in the recovery phase. There is always a risk of something going wrong and needing revision surgeries. Creating a method where both the patient and physicians have a better understanding of the probability of a revision surgery could therefore be of great interest. Using information on similar patients can create great insight into the probabilities of revisions. Grouping similar patients can function as a means to predict the outcome of a given patient. The goal of this thesis is therefore to use digital twins, a relatively new term in the medical sector, to create a method for outcome analysis. Through clinical data, we group the patients into clusters using unsupervised learning to gain insight into the characteristics and outcomes of patients. We had no prior knowledge about relationships between the patient baseline data and the possibilities of good clinical outcomes or a less favorable outcome as having problems relating to surgeries. Ideally, such knowledge should be available prior to any surgical treatment, and having an understanding of a particular patient’s chances to be successfully treated is beneficial and cost-effective. The digital twin would be a good basis for comparison since understanding what has happened to similar patients can shed light on the patient’s final outcomes. Thus the result of unsupervised was judged as a solid basis for defining a digital twin. Through both the quality register data in a synthetic version of the Norwegian Arthroplasty Register (NAR) and Medical Information Mart for Intensive Care IV (MIMIC IV) we have created an artifact in the form of a method to discover similarities in patients that can be a basis for a digital twin. The synthetic NAR data provides insight into the variables extracted from patient records. MIMIC IV provides insight into the diagnostic methods, represented by ICD 9 CM codes, giving a richer clinical picture. Resulting from both datasets are digital twins, MIMIC IV has given a more homogenous and better-distinguished cluster, while such a separation could not be reached in the NAR dataset.