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

Safety considerations are important in bridging the gap from foundation models to foundation intelligence, particularly in applications where errors can have harmful consequences. Selective prediction is a viable approach to enhancing safety by conveying uncertainty information and promoting human intervention when the algorithm lacks confidence. In the specific context of mortality risk prediction for critically ill patients, this study concentrated on selective prediction on the imbalanced dataset and proposed the self-adaptive mixup-augmented selective prediction (SAMASP) model. Experimental results demonstrated the SAMASP model's effectiveness in enhancing the training of the abstention term and reducing selective risk. To optimize the practical application of selective prediction models, we illustrated that the confidence of positive predictions could reasonably reflect precision levels. Furthermore, we presented an approach to integrate uncertainty analysis with model interpretation, providing an additional layer of safety assurance for model-based decision support in practice.