Early prediction of MODS interventions in the intensive care unit using machine learning

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Multiple Organ Dysfunction Syndrome (MODS) is a primary cause of mortality in critically ill patients.
• The lack of effective treatments necessitates early identification and intervention strategies for MODS.
• Dysregulated inflammatory response is a key mechanism underlying MODS development.

Purpose of the Study:

• To develop and validate machine learning models for early prediction of MODS.
• To quantify risk factors contributing to MODS using explainable AI techniques.
• To automatically recommend interventions for MODS based on predictive modeling.

Main Methods:

• Utilized machine learning algorithms, including stacked ensembles, for early MODS risk assessment.
• Employed Kernel SHapley Additive exPlanations (Kernel-SHAP) for quantifying prediction factors.
• Trained and tested models on MIMIC-III and MIMIC-IV databases using patient vital signs, lab results, and ventilator data.

Main Results:

• The SuperLearner model demonstrated superior performance with a Yordon index of 0.813 and accuracy of 0.893 on the MIMIC-IV test set.
• The Deep-Wide Neural Network (DWNN) achieved a maximum area under the curve of 0.960 and specificity of 0.935.
• Kernel-SHAP identified minimum Glasgow Coma Scale (GCS) and maximum MODS scores related to GCS and creatinine as significant risk factors.

Conclusions:

• Machine learning-based early warning models for MODS possess significant clinical utility.
• The SuperLearner model outperforms other common machine learning models in prediction efficiency.
• Integrating diverse counterfactual explanations (DiCE) with Kernel-SHAP facilitates practical application of automated MODS early intervention.