Predicting medical events and ICU requirements using a multimodal multiobjective transformer network.

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Critical Care Medicine

Background:

• Effective management of intensive care unit (ICU) resources is crucial for high-quality, cost-effective patient care.
• Resource non-availability can lead to adverse outcomes for critically ill patients.
• Predictive technologies offer a solution for optimizing resource utilization and patient management.

Purpose of the Study:

• To develop and validate predictive models for forecasting ICU length of stay and critical interventions.
• To leverage early patient data (within 24 hours) for accurate predictions.
• To compare the performance of a multiobjective model against independent prediction models.

Main Methods:

• Utilized the publicly available Medical Information Mart for Intensive Care (MIMIC-III v1.4) dataset for model development and cross-validation.
• Incorporated patient vital signs, laboratory measurements, and nursing notes from the first 24 hours of ICU stay.
• Employed Local Interpretable Model-agnostic Explanations (LIME) to identify key predictive features.

Main Results:

• The proposed model significantly outperformed previous models primarily relying on patient vitals for predicting ICU stay.
• A multiobjective model demonstrated superior performance compared to independent models for predicting both ICU stay and critical interventions.
• LIME successfully identified features driving prediction decisions, enhancing model interpretability.

Conclusions:

• Predictive modeling using early ICU data is effective for forecasting patient trajectories and optimizing resource management.
• Multiobjective models offer advantages over independent models for complex critical care predictions.
• Interpretable AI methods like LIME are essential for building clinical trust in predictive healthcare technologies.