Enhancing sepsis detection using feed-forward neural networks with hyperparameter tuning techniques

Abstract

This paper investigates the use of feed-forward neural networks for sepsis detection, emphasizing class imbalance mitigation and hyperparameter optimization. Leveraging random oversampling, synthetic minority over-sampling technique (SMOTE), and random sampling techniques, we address class imbalance, significantly improving feed-forward neural network performance. The resulting model achieves an impressive 83% accuracy on the test set, with notable enhancements in precision, recall, and F1-score for the positive class. Hyperparameter tuning using RandomizedSearchCV identifies optimal parameters, including an alpha value of 0.01 and the logistic activation function, leading to a remarkable 57.5% test accuracy. GridSearchCV also contributes to model refinement, albeit with a slightly lower test accuracy of 51.5%. These findings underscore the importance of robust hyperparameter tuning methods in optimizing feed-forward neural network models for imbalanced datasets, particularly in sepsis detection. The insights gained hold promise for the development of more accurate diagnostic tools, ultimately improving patient outcomes in clinical practice.