Prediction of ICU Mortality Risk Based on Task-Aware Feature Reconstruction

Predicting the mortality risk of ICU patients contributes to improving the quality of medical diagnosis and treatment, as well as optimizing resource allocation in healthcare institutions. Existing deep learning models for mortality risk prediction commonly suffer from insufficient perception of task-relevant features. In this paper, a method based on task-aware feature reconstruction is proposed for ICU mortality risk prediction. Transformer module is utilized to mask and reconstruct crucial timestamps in the original ICU sequence, extracting robust feature representations highly correlated with the prediction task. At the same time, multi-feature enhancement is carried out based on statistical features, domain transformation features and approximate entropy features, so as to improve the accuracy of ICU mortality risk prediction. Experimental validation on the MIMIC-III ICU patient dataset demonstrates that our model outperforms other models in terms of predictive capability, and provides more reliable and interpretable feature representations for mortality risk prediction tasks, offering valuable auxiliary information for clinical decision-making.