Research map / Time series / Long-term forecasting
Knowledge-Empowered Dynamic Graph Network for Irregularly Sampled Medical Time Series
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KEDGN, a graph convolutional recurrent network whose variable-specific parameters and dynamic variable graphs are derived from BERT embeddings of textual medical descriptions of each variable, with a density-aware edge adjustment, for classifying irregularly sampled medical time series.
Problem
Irregularly sampled medical time series (ISMTS) have variables with distinct temporal patterns and medical correlations that change with sampling density; existing methods fail to capture fine-grained variable-level features and learn variable graphs only from misaligned observations.
Approach
Use PLM embeddings of each variable's textual medical description to generate variable-specific GRU parameters (via a weight pool) and a static medical-knowledge variable graph, then modulate edges per timestamp with a sampling-density-aware mechanism inside a graph convolutional GRU.
Key results
- KEDGN achieves best AUROC/AUPRC on all four datasets, e.g., P19 AUROC 92.3 and AUPRC 62.5 (KEDGN-Name) vs. Warpformer 91.8/60.6.
- Outperforms strongest baseline by an average of 0.9% AUROC and 1.6% AUPRC.
- Under leave-variables-out (10-50% discarded on MIMIC-III), KEDGN attains best results on 35 of 40 metrics across the four datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KEDGN-ChatGPT instead of Warpformer | Irregular medical time series classification · P19 | AUROC: 91.8 ± 0.4 -> 92.2 ± 0.5 | Helped |
| Use KEDGN-ChatGPT instead of Raindrop | Irregular medical time series classification · P19 | AUPRC: 61.2 ± 1.1 -> 62.0 ± 1.3 | Helped |
| Use KEDGN-ChatGPT instead of mTAND | Irregular medical time series classification · Physionet | AUROC: 86.9 ± 1.3 -> 87.9 ± 0.2 | Helped |
| Use KEDGN-ChatGPT instead of IP-Net | Irregular medical time series classification · Physionet | AUPRC: 55.8 ± 1.4 -> 57.1 ± 1.8 | Helped |
| Use KEDGN-ChatGPT instead of Warpformer | Irregular medical time series classification · MIMIC-III | AUROC: 84.6 ± 0.5 -> 85.1 ± 0.3 | Helped |
| Use KEDGN-ChatGPT instead of Warpformer | Irregular medical time series classification · MIMIC-III | AUPRC: 47.4 ± 0.9 -> 48.3 ± 1.6 | Helped |
| Use KEDGN-ChatGPT instead of StraTS | Irregular medical time series classification · P12 | AUROC: 86.7 ± 0.7 -> 87.8 ± 0.5 | Helped |
| Use KEDGN-ChatGPT instead of StraTS | Irregular medical time series classification · P12 | AUPRC: 52.1 ± 1.5 -> 54.5 ± 1.5 | Helped |
| Add variable-specific RNN parameter weights (ablation) | KEDGN · P19 | AUROC: 91.5 ± 0.3 -> 92.2 ± 0.5 | Helped |
| Use textual variable representations instead of random learnable embeddings (ablation) | KEDGN · P19 | AUROC: 91.7 ± 0.6 -> 92.2 ± 0.5 | Helped |
| Model inter-variable dependencies with the dynamic graph (ablation) | KEDGN · P19 | AUROC: 90.8 ± 0.8 -> 92.2 ± 0.5 | Helped |
| Use knowledge-empowered node embeddings instead of random learnable embeddings (ablation) | KEDGN · P19 | AUROC: 91.5 ± 0.6 -> 92.2 ± 0.5 | Helped |
Surprising findings
- Replacing text embeddings with randomly initialized learnable variable embeddings lowers performance, while removing the graph hurts less on some datasets, and the default BERT text embedding slightly hurts P12 relative to no-text variant.
- Decoder-only GPT2 embeddings performed notably worse on P19 and MIMIC-III than encoder models.
Limitations
- Recurrent backbone has sequential computation that limits runtime.
- Method is tailored to medical settings where variables have textual domain knowledge; may not transfer to e.g. human activity recognition.
- Runtime is 1.3x that of Warpformer on Physionet.
Experimental setup
Datasets: P19 P12 MIMIC-III Physionet
Benchmarks: P19 Physionet MIMIC-III P12
Metrics: AUROC AUPRC
Baselines: GRU-D ODE-RNN IP-Net SeFT mTAND Raindrop StraTS DuETT ViTST Warpformer DGM2-O MTGNN
Builds on: GCRNN GRU-D Raindrop Warpformer MTGNN
Closest papers by meaning
- MIRA: Medical Time Series Foundation Model for Real-World Health Data (NeurIPS 2025 · similarity 0.85)
- From Indicators to Insights: Diversity-Optimized for Medical Series-Text Decoding via LLMs (NeurIPS 2025 · similarity 0.85)
- JoLT: Jointly Learned Representations of Language and Time-Series for Clinical Time-Series Interpretation (Student Abstract) (AAAI 2024 · similarity 0.84)
- Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval (ICLR 2025 · similarity 0.83)
- Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach (ICML 2024 · similarity 0.83)
- GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.