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

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Clinical Decision Support Systems

Background:

• Intensive care units (ICUs) require timely and accurate clinical decisions regarding medication administration and dosage.
• Existing decision support tools may not fully capture the dynamic nature of patient treatment over time.
• Reducing clinician cognitive load is crucial for improving patient safety and outcomes.

Purpose of the Study:

• To develop and evaluate a time-oriented machine learning tool for predicting medication administration decisions and dosages.
• To enhance the accuracy and reliability of automated clinical decision support systems.
• To improve the prediction of treatment dynamics in critical care settings.

Main Methods:

• Utilized a Long Short-Term Memory (LSTM) based neural network architecture.
• Implemented a partitioned prediction horizon (12-hour window into three sub-windows) to better model treatment dynamics.
• Introduced a sequential prediction process: a binary treatment-decision model followed by a quantitative dose-decision model.
• Incorporated non-temporal features (e.g., patient age) into the temporal network using two distinct methods.

Main Results:

• The developed LSTM model demonstrated improved performance in predicting medication decisions and dosages on the MIMIC-IV ICU database.
• Partitioning the prediction horizon and employing a sequential prediction process led to enhanced accuracy.
• The inclusion of non-temporal features further refined the predictive capabilities of the model.

Conclusions:

• The time-oriented machine learning tool offers a promising advancement in clinical decision support for medication management.
• The model's ability to predict treatment dynamics and incorporate diverse features contributes to a more reliable decision-support system.
• This approach has the potential to significantly reduce clinicians' cognitive load and support evidence-based treatment decisions.