CT-HEG: A Novel Graph Schema for ICU Mortality Prediction

A new schema called CT-HEG (Continuous-Time Heterogeneous EHR Graph) has been developed by researchers to model irregular clinical observations in ICU environments. This schema represents each ICU admission as a typed, timestamped graph featuring three types of nodes (visit, vital, lab_event) and 2D edge attributes that reflect timing and value without the need for imputation. The researchers implemented CT-HEG as CHIRP-Net, a four-layer heterogeneous GATv2Conv network, and tested it on MIMIC-IV v3.1, which includes 31,142 ICU stays lasting at least 48 hours, with a mortality rate of 13.4%. CHIRP-Net achieved a 5-seed mean AUROC of 0.8449±0.0000, surpassing baseline models such as logistic regression, mTAND, a Transformer, and GRU-D. An ablation study was performed to identify the architectural choices that enhance predictive performance. This research addresses the shortcomings of current sequence models that manage irregular sampling but overlook typed relational structures, as well as graph models that presume fixed-interval inputs. The findings are accessible on arXiv with the identifier 2608.02663.