Out-of-hospital cardiac arrest is a major health problem. Outcome varies, stressing the need for improvement and coherence to uniform reporting. With the advantages of artificial intelligence, this study aims to improve the Danish cardiac arrest registry. We extract features from the text and apply them to machine learning models. Thus, we evaluate characteristics of cardiac arrest. This may substitute the annual manual validation and it proposes improvement and development of health research.
Using artificial intelligence to improve the Danish Cardiac Arrest Registry
Out-of-hospital cardiac arrest (OHCA) is a leading cause of death worldwide and patient outcome vary substantially throughout regions suggesting additional evaluation and potential for improvement. When focussing on subgroups of OHCA of non-cardiac origin, data remain scarce and the need of revised guidelines is evident. With aid of machine learning, the hypothesis is that advances in text searches may lead to improvement of quality of data from the Danish Cardiac Arrest Registry.
This project enables the assessment of novel information regarding subgroups of OHCA. The way of using artificial intelligence within text mining allows high quality use of data in order to strengthen the results and the Danish Cardiac Arrest Registry. Throughout analyses, a better understanding of the preceding circumstances, etiology and prehospital assessment might contribute to improve handling these types of arrests. Further, the applied methodology may substitute the manual validation of the around 9000 cases per year in Denmark. Finally, it proposes improvement of quality and development of observational health research.
Data will be evaluated using a bag-of-words representation which enables representation of documents as vectors. In this case, the documents are comprised of the electronic prehospital patient report. This approach allows classification using supervised machine learning algorithms such as naïve bayes, support vector machine, ensemble methods and random forest. Word embeddings and a latent dirichled allocation approach will be applied to compare unsupervised learning based on clustering with the above-mentioned supervised classifiers.
Professor Christian Torp-Pedersen, DMSc, MD, Aalborg University Department of Health Science and Technology, Nordsjællands Hospital, Department of Cardiology
Professor Søren Mikkelsen. MD, PhD, Prehospital Research Unit, the Region of Southern Denmark & Odense University Hospital
Stig Nikolaj Fasmer Blomberg, MD, PhD, Copenhagen Emergency Medical Services, Copenhagen University Hospital