Predicting Hemodynamic and Pulmonary Decompensation with Deep Neural Networks: Performance and Explainability.

Area of Science:

• * Artificial Intelligence in Medicine
• * Critical Care Medicine
• * Machine Learning for Healthcare

Background:

• * Predicting hemodynamic and pulmonary decompensation in intensive care units (ICUs) is challenging for medical staff.
• * Artificial intelligence (AI) offers potential support for physicians in critical care safety.
• * Early detection of decompensation is vital to prevent severe outcomes like organ damage or death.

Purpose of the Study:

• * To develop and evaluate Transformer networks for predicting patient decompensation scores.
• * To forecast hemodynamic and pulmonary decompensation up to 24 hours in advance.
• * To enhance explainability of AI models using feature attribution techniques.

Main Methods:

• * Application of Transformer networks for time-series prediction of patient states.
• * Utilization of Shapley values for analyzing feature attribution and model interpretability.
• * Development of novel techniques for aggregating and visualizing feature attribution in time-series data.

Main Results:

• * Transformer networks accurately predicted hemodynamic decompensation with a mean error of 3.73%.
• * Pulmonary decompensation was predicted with a mean error of 2.2%.
• * Feature attribution analysis revealed key medical variables influencing predictions and their timing.

Conclusions:

• * AI-driven prediction models can significantly support clinical decision-making in ICUs.
• * Explainable AI methods provide insights into the factors driving decompensation predictions.
• * Timely prediction of decompensation is crucial for effective medical intervention and improved patient outcomes.