Machine Learning Based Early Detection System of Cardiac Arrest
Journal
Proceedings - 2019 International Conference on Technologies and Applications of Artificial Intelligence, TAAI 2019
Pages
8959922
Date Issued
2019
Author(s)
Abstract
Target - Most of the Cardiac Arrest (CA) cases are preventable because the CA patients usually had abnormal clinical signs or symptoms before their suffering from CA. In general, the appropriate steps of Cardiopulmonary resuscitation (CPR) for CA patients will increase the patients' survival rate and reduce the consequent medical expenses. Accordingly, we propose a system and the related methods for detecting CA before the CPR event occurred earlier and it is not only assisting physicians to early diagnose of CA and immediately warning but also improving the medical quality.Methods - In this study, the raw dataset is collected from the electronic health records (EHRs) of the adult patients (age ≧20 years) who visited emergency department (ED) and stayed in the emergency detention area for more than 6 hours during January 2014 to December 2015, and it is provided by National Taiwan University Hospital (NTUH). We perform the data preprocessing and cleaning for the dataset using a resampling technique to balance the data amount of CPR and Non-CPR patients of the dataset, and then we construct a sliding window and apply several classifiers for model training and reducing the possible overfitting problem. Additionally, we use the measures such as the Area Under the Receiver Operating Characteristic Curve (AUROC) and the Area Under the Precision-Recall Curve (AUPRC) to comparative evaluate the performance of our models built.Results - Our approach avoids the problems of dataset imbalance and possible overfitting effectively. The performance among classifiers selected show that the best one is random forest (RF) when CPR event happened, but the better ones are Logistic Regression and LSTM than the remaining classifiers and close to that of RF during 1 to 4 hours before the CPR time.Conclusion - We have the contribution in predicting CA and CPR event before it occurred around 3 to 3.5 hours in advance, and it is similar to that of the state-of-the-art such Early Warning Score (EWS) with around 3.5 hours. In addition, avoiding the problem of dataset imbalanced may effectively improve the accuracy of predicting CA as well. Accordingly, it helps to assist the emergency clinical physicians to achieve the hospital's quality management including the clinical or medical resources allocation.
Type
conference paper