(1353) Augmenting ICU Mortality Prediction in Pulmonary Embolism Patients Using Imbalanced Data Techniques

Introduction: Imbalance in classes of an outcome variable can cause difficulties for machine learning algorithms, leading to biased results towards the majority class. Such is the case in the eICU dataset, where ICU mortality among Pulmonary Embolism (PE) patients has a severe imbalance ratio of 0.0457, indicating more survival than death events. This study aims to demonstrate that advanced imbalanced data techniques can improve prediction of severely imbalanced outcomes.

Methods: From an initial set of 25 features, we used the SelectFromModel feature selection technique with LightGBM (LGBM) classifier, to choose the top 19 features. The selection was made based on the feature importance in predicting ICU mortality. Machine learning algorithms struggle with imbalanced data as they inherently assume an equal distribution of classes. To overcome this, we used the Synthetic Minority Over-sampling Technique (SMOTE), which creates synthetic minority class instances by interpolating between existing ones, thus balancing the class distribution.

Results: Using the Matthews Correlation Coefficient (MCC) as our primary metric, a suitable choice given the class imbalance, the LGBM model achieved an MCC of 0.3205, an F1 score of 0.2667, and a recall of 0.1667. Application of SMOTE improved the MCC to 0.3635, with a corresponding F1 score of 0.3871, and a recall of 0.5, indicating enhanced prediction of mortality in the imbalanced dataset.

Conclusions: Our findings support the use of advanced imbalanced data techniques, specifically SMOTE, in improving the performance of machine learning models in scenarios of severe outcome imbalance. Future work should explore the performance of these methods across different healthcare datasets with varying degrees of imbalance.