Time-aware Adversarial Networks for Adapting Disease Progression Modeling

Modeling patient disease progression using their historical Electronic Health Records (EHRs) is critical to assist clinical decision making and provide prompt medications. Long-Short Term Memory (LSTM) has been widely applied to handle large-scale sequential data such as EHRs. However, when applied to EHRs, the standard LSTM encounters one major limitation: lack of the generalizability given the intricate heterogeneity among different patient groups. To overcome this limitation, we propose a domain adaptation (DA) framework named Time-aware Adversarial Networks, which combines two temporal LSTM-based recurrent components, a label predictor and a domain classifier, to extract an invariant feature representation across different patient groups (domains) through an adversarial learning process. The label predictor learns the temporal developing patterns of the disease while the domain classifier identifies the temporal changes of feature distributions across patient groups. We evaluate the proposed DA framework on modeling the progression of an extremely challenging disease, septic shock, by using real-world EHRs among different patient groups across age, race, and gender. Our results demonstrate that our proposed DA framework can significantly improve the performance for the disease progression modeling while standard machine learning models are severely impacted by the data heterogeneity in EHRs.