Predicting ICU Interventions: A Transparent Decision Support Model Based on Multivariate Time Series Graph...

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Clinical Decision Support

Background:

• Intensive care unit (ICU) patient management requires timely interventions based on complex, dynamic data.
• Existing prediction models often lack interpretability, hindering clinical trust and adoption.
• Accurate prediction of interventions like mechanical ventilation and vasopressors is crucial for patient outcomes.

Purpose of the Study:

• To develop and validate a novel multivariate time series graph convolutional neural network (GCN) for predicting ICU interventions.
• To enhance the interpretability of AI-driven predictions in critical care settings.
• To improve the accuracy and timeliness of clinical decision-making for ICU patients.

Main Methods:

• Utilized the MIMIC-III database containing real-world ICU patient records.
• Developed a multivariate time series GCN model integrating physiological signals, drug data, and patient characteristics.
• Performed adjacency matrix importance analysis to assess model interpretability and feature relevance.

Main Results:

• Achieved significant improvements in predicting mechanical ventilation: accuracy increased from 81.6% to 91.9%, F1 score from 0.524 to 0.606.
• Demonstrated enhanced prediction for vasopressor interventions: accuracy rose from 76.3% to 82.7%, F1 score from 0.509 to 0.619.
• Interpretability analysis confirmed the model's reliance on clinically meaningful features.

Conclusions:

• The proposed GCN model offers a powerful, interpretable tool for predicting critical ICU interventions.
• This approach significantly outperforms existing methods, enhancing accuracy and providing clinical insights.
• The study advances AI-driven decision support systems, promising improved patient safety and outcomes in ICUs.