CT-HEG: A Bidirectional, Timestamp-Attributed Event Graph for ICU In-Hospital Mortality Prediction - An Architectural Ablation Study
CT-HEG introduces a novel schema encoding ICU stays as typed, timestamped heterogeneous graphs with three node types (visit, vital, lab_event) and 2D edge attributes capturing timing and normalized values without imputation CHIRP-Net, a four-layer heterogeneous GATv2Conv instantiation, achieved AUROC 0.8449±0.0071 on MIMIC-IV v3.1 (31,142 ICU stays, 13.4% mortality), with an ensemble reaching 0.8618 Bidirectional connectivity proved essential: removing reverse edges collapsed AUROC by 0.1968±0.0
Analysis
TL;DR
- CT-HEG introduces a novel schema encoding ICU stays as typed, timestamped heterogeneous graphs with three node types (visit, vital, lab_event) and 2D edge attributes capturing timing and normalized values without imputation
- CHIRP-Net, a four-layer heterogeneous GATv2Conv instantiation, achieved AUROC 0.8449±0.0071 on MIMIC-IV v3.1 (31,142 ICU stays, 13.4% mortality), with an ensemble reaching 0.8618
- Bidirectional connectivity proved essential: removing reverse edges collapsed AUROC by 0.1968±0.0073, disconnecting observation nodes from the visit readout
- Surprisingly, collapsing all heterogeneous edge types into a single relation (7× fewer parameters) outperformed the full model across all seeds, suggesting edge-type specialization adds little predictive value
- The model achieved good calibration (ECE 0.0307 after temperature scaling), but external validation, temporal evaluation, and demographic fairness audits remain necessary before clinical deployment claims
Why It Matters
This work bridges a critical gap between sequence-based and graph-based approaches for irregular EHR data, demonstrating that typed relational structure matters less than bidirectional connectivity and temporal edge attributes for mortality prediction. For AI practitioners building clinical prediction systems, it provides a practical, well-calibrated architecture with open-source code and a rigorous ablation study that clarifies which design choices actually drive performance.
Technical Details
- CT-HEG Schema: Each ICU stay is represented as a directed, timestamped heterogeneous graph with node types
visit,vital, andlab_event. Edges carry 2D attributes(t_hours/48, value_norm)encoding time since admission (normalized to 48-hour window) and normalized clinical values, eliminating the need for imputation or fixed-interval binning. - CHIRP-Net Architecture: A four-layer heterogeneous Graph Attention Network using GATv2Conv layers, with separate attention mechanisms per edge type. The model aggregates observation-level representations through bidirectional edges into a visit-level readout for mortality classification.
- Evaluation Setup: Tested on MIMIC-IV v3.1 with 31,142 ICU stays (length of stay ≥48h, 13.4% mortality rate). Five random seeds with bootstrapped confidence intervals. Baselines include logistic regression, mTAND, a standard Transformer, and GRU-D.
- Ablation Findings: Reverse edges contributed the largest performance gain (ΔAUROC ≈ 0.197). Time-attentive edge features added ~0.025 AUROC. Edge-type collapsing (unifying all relations) improved performance while reducing parameters 7×. Post-calibration expected calibration error was 0.0307.
Industry Insight
- The counterintuitive finding that edge-type collapsing outperforms the full heterogeneous model suggests practitioners should prioritize simpler, more parameter-efficient graph designs over complex typed-edge architectures in clinical EHR settings, potentially saving significant compute without sacrificing accuracy.
- Bidirectional graph connectivity should be considered a non-negotiable design choice for any graph-based clinical prediction system; unidirectional architectures risk severing the information flow from observations to outcome predictions entirely.
- While the model shows strong calibration—a critical requirement for clinical deployment—the authors correctly flag that external validation, prospective temporal evaluation, and fairness auditing are prerequisites before any real-world clinical adoption claims, setting a responsible benchmark for the medical AI community.
Disclaimer: The above content is generated by AI and is for reference only.