hotwheels

This project develops a domain-adaptive Transformer framework for robust ICU mortality prediction across heterogeneous hospital datasets. We begin with a pretrained eICU Transformer model and introduce a domain-adversarial adaptation module that aligns its latent representations with those of MIMIC-IV using a Gradient Reversal Layer and a lightweight domain discriminator. This allows the model to preserve clinically meaningful mortality-predictive structure learned from eICU while learning to ignore dataset-specific artifacts such as differing measurement frequencies, variable distributions, and documentation patterns. By combining labeled eICU data with unlabeled or limited-label MIMIC data, the system learns domain-invariant patient embeddings that improve cross-center generalization and reduce performance degradation when models are deployed in new hospitals. This work provides a practical, modern baseline for multi-center ICU risk modeling and demonstrates the value of adversarial domain adaptation for clinical deep learning.

Built With

• python
• pytorch