Abstract P164: Machine Learning in Predicting Survival from a Web-based Registry of In-hospital Resuscitation: Implication for Reassessing Prognostic Determinants in Utstein Template
Introduction: Little is known about machine learning (ML) performance in survival prediction from in-hospital resuscitation.
Objectives: To compare performances of ML and logistic regression (LR) in survival prediction from in-hospital resuscitation and to assess prognostic determinants selected from different strategies.
Methods: In a tertiary teaching hospital, a total of 1048 adults (≥18 years) events were assessed from the Web-Based Registry System on In-hospital Resuscitation (WRSIR) between 1 January 2005 and 31 December 2007. The main outcome was survival to discharge. Utstein variables and anticipated viewpoints about resuscitation events from the related doctor and family members were collected prospectively. Samples were randomly divided into the training and testing datasets with a 3-to-1 ratio. Expert opinions, univariate analysis, and entropy measures were used to select variables. LR and three ML models including decision tree, artificial neural networks (ANN), and k-nearest neighbor (kNN) were assessed. Performances were evaluated with area under the ROC curve (AUC) in the testing dataset and significances were examined between models.
Results: The rate of survival to discharge was 17%. Ten factors were chose by LR. Eight more factors (arrest location, immediate cause arrhythmias, first monitored rhythm, causes of arrest cardiac diseases, airway before intubation, immediate cause hypotension, massage attempt, and anticipated by family) were identified by ML approaches with improved AUC when massage duration was discretized by 15 minutes. Overall there were no differences in AUC between models (Table⇓).
Conclusions: ML approaches can provide comparable performance as LR in predicting survival following in-hospital resuscitation and show convincing evidence of future Utstein variables selection strategy. Furthermore, massage duration of more than 15 min can be used as poor prognostic factor to help end-of-life decision making.