A Novel Multi-Task Teacher-Student Architecture With Self-Supervised Pretraining for 48-Hour Vasoactive-Inotropic...

Area of Science:

• Critical Care Medicine
• Artificial Intelligence
• Health Informatics

Background:

• Sepsis is a leading cause of intensive care unit (ICU) mortality, necessitating early detection and intervention.
• Predicting sepsis is challenging due to dynamic vasoactive-inotropic score (VIS) variations, data irregularities, and confounding factors.

Purpose of the Study:

• To develop a novel Teacher-Student multitask framework with self-supervised Masked Autoencoder (MAE) pretraining for improved sepsis prediction.
• To enhance the adaptation of time-series representations to heterogeneous VIS data for more robust sepsis identification.

Main Methods:

• Implemented a Teacher-Student multitask framework utilizing self-supervised VIS pretraining via MAE.
• The teacher model conducted mortality classification and severity-score regression.
• The student model distilled robust time-series representations for heterogeneous VIS data adaptation.

Main Results:

• Achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.829 on MIMIC-IV 3.0, outperforming baseline LSTM methods (0.74).
• SHAP analysis identified SOFA score, LODS, marital status, Medicaid insurance, and SAPSII as significant predictors of ICU mortality.
• Highlighted the influence of sociodemographic factors alongside clinical scores in predicting ICU mortality.

Conclusions:

• The proposed AI framework enables earlier identification of high-risk sepsis patients, improving prediction accuracy.
• Integrating clinical and sociodemographic factors enhances ICU decision-making and patient management.
• The developed multitask and distillation strategies offer advanced tools for ICU decision support and sepsis management.