Graph-Enhanced Disease Trajectory Discovery and Temporal Endpoint Forecasting from Longitudinal Electronic Health Records
A MIMIC-IV study finds graph branching motifs mainly improve long-range event timing, not endpoint discrimination.
Researchers describe a leakage-controlled, graph-enhanced framework on MIMIC-IV v3.1 that learns train-only admission transitions and freezes endpoint-specific branching motifs. A catalog of 99,527 motifs is retrieved, at most 24 per endpoint, by a time-aware Transformer. Across five seeds, micro-AUPRC rose from 0.9330 to 0.9357 and event-time MAE fell from 676.9 to 657.7 days, with the timing gain concentrated on events beyond 730 days. Corrected survival metrics did not show superiority over a matched history-only discrete-hazard model, and swapping in GNN edge scores did not improve forecasting.
- Patient-split MIMIC-IV v3.1 graph yields 99,527 branching motifs.
- Micro-AUPRC rose from 0.9330 to 0.9357 across five seeds.
- Event-time MAE fell 19.2 days, mainly for events after 730 days.
- Survival metrics did not beat a matched history-only hazard model.
- GNN edge repair did not improve downstream forecasting.
From the original article
1 min read at arxiv.orgLongitudinal electronic health records (EHRs) can reveal disease-progression structure, while useful forecasting requires leakage-controlled discovery, censor-aware outcomes, and inspectable trajectory evidence. We developed a graph-enhanced framework in MIMIC-IV v3.
By Jin-Jia Zheng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.