DeepSelective: Interpretable prognosis prediction via feature selection and compression in EHR data

01 / Research question

Can an EHR model preserve predictive information while showing which measured features matter?

02 / Innovation

DGFS learns a sparse feature mask; ATA compresses the selected-feature input, and RML combines complementary representations before prognosis prediction.

03 / Mechanism

Gumbel-Softmax gates select features; a transformer autoencoder encodes the masked input. Representation Matching combines aligned and complementary representations. Prediction, reconstruction and alignment losses train the network jointly; a PID-inspired controller adjusts sparsity.

04 / Experimental setup

Manuscript v2 follows the referenced MIMIC-III preprocessing and partitioning protocol for mortality and decompensation. Baselines include RETAIN, T-LSTM, SAnD, ConCare and AdaCare. Exact cohort counts and split proportions are not restated in §5.2. The journal version adds MIMIC-IV; these values remain v2 results.

Author manuscript v2 · 20 June 2025 · Table 2 · Source ↗

05 / Results

Table 2 reports mortality AUROC 0.9054 and decompensation AUROC 0.9143; selected baseline comparisons appear below.

06 / Ablation & limitations

Without ATA, mortality AUROC is 0.8524 and AUPRC 0.4781 (Figure 5). Selection improves inspectability but does not prove causal biomarkers. The masking argument concerns access to unselected features, not train/test leakage. External hospitals, missingness and controller stability require further validation.

The displayed MIMIC-III results belong to author manuscript v2. Final journal evaluation also includes MIMIC-IV; these values must not be transferred across versions.

07 / My contribution

Co-first author with Qian Yang. The final contribution statement credits methodology, investigation, data curation, project administration, visualization and writing.

08 / Paper & code