Treatment Prediction in the ICU Setting Using a Partitioned, Sequential Deep Time Series Analysis

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Machine Learning for Healthcare

Background:

• Accurate patient state evaluation is crucial for effective treatment.
• Predicting medication treatment decisions and dosages presents significant challenges.
• Existing models struggle with classifying 'no treatment' scenarios.

Purpose of the Study:

• To develop a novel neural network architecture for patient state evaluation.
• To predict binary medication treatment decisions and specific dosages.
• To address the challenge of 'no treatment' classification in dose-prediction models.

Main Methods:

• Utilized temporal data, patient demographics, and comorbidities.
• Developed a neural network architecture for state evaluation and prediction.
• Partitioned a 12-hour window into sub-windows for enhanced training data utilization.
• Employed sequential prediction to enable 'no treatment' classification.

Main Results:

• The model demonstrated ability to predict treatment decisions and dosages in hypokalemia, hypoglycemia, and hypotension.
• Partitioned analysis improved performance by utilizing data from previous sub-windows.
• Sequential prediction successfully addressed the 'no treatment' classification issue.

Conclusions:

• The developed neural network offers a robust approach to patient treatment prediction.
• The partitioned analysis and sequential prediction methods enhance model performance and applicability.
• This AI-driven approach has the potential to optimize clinical decision-making in critical care settings.