Benchmarking LSTM, Lightgbm, And Xgboost For Enhanced Cardiac Arrest Predictive Performance

Abstract

Since cardiac arrest is a fatal medical emergency with a high death rate, it's critical to find efficient ways to anticipate it and take action as soon as possible. Despite the shortcomings of conventional risk assessment methods, ML is a rapidly emerging technology that promises to analyse complex clinical data to identify patients who are at risk.

To address this problem, we choose and compare three state-of-the-art ML models in this paper: LSTM, LightGBM, and XGBoost. These will be the main focus for predicting the cardiac arrest episodes with structured patient data.

During this process both models are trained and tested on the whole data set, and are assessed based on five metrics: accuracy, precision, recall, F1 score, and AUROC. In our comparison, we were able to find out what each model offers. The overall accuracy and AUROC scores were best for LightGBM. In a clinical environment, it is important for the algorithm to have the lowest false positive rate among them, which in this case was XGBoost. LSTM's highest recall score, on the other hand, shows that it identifies the most authentic positive cases. Results indicate that none of the models is superior for all. Instead, to make clinical decision support systems more reliable, a multi-model approach such as the ensemble technique could take advantage of the different strengths of these models. This study's findings add to the body of research and offers valuable data for future research and application in the healthcare field.