Tachycardia Prediction With LSTM And XGBoost : A Comparative Analysis of Machine Learning Models

Sammanfattning

This paper presents the implementation and evaluation of multiple LSTM-based and XGBoost-based models for the prediction of tachycardia. The models were trained on a dataset comprising 6,168 labeled cases, including 303 tachycardia episodes and 5,883 non-tachycardia episodes. The LSTM models achieved AUC scores ranging from 0.5284 to 0.6086, with corresponding F1 scores between 0.0959 and 0.1202. In comparison, the XGBoost models performed better, obtaining AUC scores between 0.5805 and 0.7015 and F1 scores between 0.1362 and 0.1778. While XGBoost demonstrated improved discriminative ability over the LSTM approaches, overall performance across all models remained limited. These results suggest that further improvements in model performance would likely require enhanced hyperparameter optimization, larger data sample and potentially longer or recording windows. Despite this, the findings indicate that traditional rule-based methods remain more reliable than the evaluated machine learning approaches for this task under the current constraints.