Sampling methods and feature selection for mortality prediction with neural networks.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning in healthcare

Background:

• The increasing digitization of healthcare data fuels the demand for automated decision support systems.
• Mortality prediction is a critical component of clinical decision-making and patient management.
• Sophisticated machine learning models, particularly Artificial Neural Networks (ANNs), offer promising performance for complex predictive tasks.

Purpose of the Study:

• To evaluate the reproducibility of a previously published mortality prediction approach utilizing ANNs.
• To assess the generalizability of the ANN model to a larger, more diverse dataset.
• To develop and optimize an ANN model for accurate in-hospital mortality prediction.

Main Methods:

• Development of an extensive data preprocessing pipeline.
• Evaluation of various data sampling techniques to handle class imbalance.
• Exploration and comparison of different Artificial Neural Network architectures.
• Training the ANN model using a loss function optimizing both precision and recall.
• Feature engineering to enhance predictive power.

Main Results:

• The optimized ANN model achieved an accuracy, sensitivity, and Area Under the Receiver Operating Characteristic (AUROC) score exceeding 0.8.
• The study successfully reproduced and generalized the mortality prediction approach.
• The preprocessing pipeline and hyperparameter tuning contributed to improved model performance.

Conclusions:

• Artificial Neural Networks are effective tools for predicting in-hospital mortality with high accuracy.
• Reproducibility and generalizability of machine learning models are crucial for reliable clinical application.
• The developed methodology provides a robust framework for data-driven mortality prediction in healthcare settings.