A Self-Adaptive Mixup-Augmented Selective Prediction Framework: A Case Study on In-Hospital Mortality Prediction

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Clinical Decision Support

Background:

• Foundation models require safety enhancements for real-world applications, especially where errors have severe consequences.
• Selective prediction offers a method to manage algorithmic uncertainty and prompt human intervention when confidence is low.
• Imbalanced datasets in critical care pose challenges for accurate mortality risk prediction.

Purpose of the Study:

• To develop and evaluate a selective prediction model for imbalanced mortality risk data in critically ill patients.
• To improve the safety and reliability of AI-driven decision support in high-stakes medical scenarios.
• To integrate uncertainty quantification with model interpretability for enhanced trust.

Main Methods:

• Proposed the self-adaptive mixup-augmented selective prediction (SAMASP) model.
• Focused on selective prediction techniques tailored for imbalanced datasets.
• Integrated uncertainty analysis with model interpretation methods.

Main Results:

• The SAMASP model demonstrated effectiveness in improving the training of the abstention term.
• The model successfully reduced selective risk in mortality prediction.
• Positive prediction confidence was shown to correlate with precision, aiding practical application.

Conclusions:

• Selective prediction, particularly with the SAMASP model, enhances safety in critical care mortality risk prediction.
• The confidence-precision relationship provides a practical metric for selective prediction models.
• Integrating uncertainty analysis with interpretation offers robust safety assurance for clinical decision support systems.