Skip to content

Deep Learning for Stroke Mortality Prediction in eICU: A Dual-Tower Transformer Framework

Zhengrong Jia*, Kwong-Cheong Wong*  ★ Best Industrial Paper ★ Best Presentation

The 6th International Conference on Computer Communication and Artificial Intelligence (CCAI), 2026

We propose a novel Dual-Tower Transformer (DT-Transformer) for stroke mortality prediction on the multicenter eICU Collaborative Research Database. The decoupled architecture processes categorical demographics and numerical vitals through separate tower pathways, each equipped with Self-Attention, before fusing representations for final prediction. The model achieves an AUPRC of 0.6171 — a 14.4% improvement over the strongest neural baseline. An Adaptive Runtime Safeguard is integrated for inference stability against physiological outliers, and attention map visualizations provide clinical interpretability.

Model AUROC AUPRC
DT-Transformer (ours) 0.8848 ± 0.0034 0.6171 ± 0.0058
XGBoost 0.8908 0.6467
Random Forest 0.8806 0.6236
Neural Network 0.8582 ± 0.0018 0.5394 ± 0.0054
Standard Transformer 0.8457 ± 0.0129 0.5279 ± 0.0195
Standard MLP 0.8534 ± 0.0058 0.5170 ± 0.0081

All reported metrics use 5-fold stratified cross-validation. XGBoost and Random Forest use single-run evaluation.

  • Dual-Tower Design — Separate encoding pathways for categorical (demographics) and numerical (vitals) features
  • Self-Attention Layers — Each tower applies multi-head Self-Attention for intra-modality feature interaction
  • Late Fusion — Tower outputs concatenated before classification head
  • Adaptive Runtime Safeguard — Detects out-of-distribution inputs at inference time for clinical safety
  • Attention Visualization — Heatmaps over input features for clinical interpretability

Stack: Python · PyTorch · Scikit-learn · XGBoost · Pandas · NumPy

eICU Collaborative Research Database — Multicenter critical care database (200,859 admissions, 208 hospitals). Stroke cohort extracted with ICD-9 codes 430–438. Features: demographics, vitals, lab values, GCS scores. Access via PhysioNet credentialed access.

Open-source implementation at github.com/ZR-JIA/Dual-Tower-Transformer-eICU-Stroke. Repository includes PyTorch model, data preprocessing pipeline, 5-fold cross-validation scripts, and attention visualization tools.

@inproceedings{jia2026dttransformer,
  title     = {Deep Learning for Stroke Mortality Prediction
               in eICU: A Dual-Tower Transformer Framework},
  author    = {Jia, Zhengrong and Wong, Kwong-Cheong},
  booktitle = {Proceedings of the 6th International Conference on
               Computer Communication and Artificial Intelligence (CCAI)},
  year      = {2026},
  note      = {Accepted Feb 2, 2026; Presented at CCAI 2026, May 24, 2026}
}