Reproduce “DeepSelective: Interpretable Prognosis Prediction via Feature Selection and Compression in EHR Data” on OpenResearch

About this paper

Ruochi Zhang, Qian Yang, Xiaoyang Wang, Tian Wang, Qiong Zhou, Ziqi Deng, Kewei Li, Yueying Wang and 5 more

Researchers from Jilin University developed DeepSelective, an end-to-end deep learning framework that predicts patient prognosis from Electronic Health Records (EHRs) by balancing predictive accuracy with model interpretability. The framework employs a dual compression strategy for feature selection and representation learning, achieving higher AUROC and AUPRC scores on MIMIC-III for both mortality and decompensation prediction compared to existing methods, while also identifying clinically significant features.

Key insights

• The dual compression strategy effectively retains informative features, with DGFS identifying statistically and clinically significant features (e.g., vital signs, GCS scores) that directly drive predictions.
• The synergistic integration of dynamic feature selection (DGFS), an attentive transformer autoencoder (ATA), and representation matching (RML) is crucial for balancing predictive accuracy with model interpretability.
• Both sparsity and perceptual compression significantly improve the separability of patient outcome classes in the latent space, indicating more discriminative representations compared to raw input features.

Summary by alphaXiv.