Time-Resolved ICU Risk Signals for Pharmacist Triage

• Across 40,856 admissions, DynaGraph outperformed 14 state-of-the-art baselines: on MIMIC-III it achieved AUROC 0.856, AUPRC 0.461, sensitivity 85.22% (+12 percentage points vs MedGNN); on eICU AUROC 0.802 (sensitivity 86.00%); on HiRID-ICU AUROC 0.881 (sensitivity 86.20%).

• For pharmacy operations, higher sensitivity combined with time-resolved feature relationships can help prioritise deterioration and organ-failure reviews; the model reports millisecond-range inference (≈13 ms) on GPU/sidecar hardware, supporting potential near-real-time surveillance but requiring appropriate runtime infrastructure and governance.

• In high-acuity care, pharmacy teams need earlier, reliable signals of deterioration to adjust therapy; static threshold scores and static models can miss evolving cross-organ patterns and provide limited actionable rationale, producing reactive workflows and manual chart review.

• Severe class imbalance and temporal instability in EHRs make rare, high-stakes events easy to miss; modelling how lab–vital relationships change over time with time-specific drivers targets these blind spots.

• Linking elevated risk to physiologic rationale can populate clinical decision support worklists that prioritise pharmacist review, strengthening stewardship and organ-failure surveillance while focusing limited resources where they matter most.

• Retrospective evaluation on 40,856 admissions from MIMIC-III, eICU (cardiac subset), HiRID, and EHRSHOT using demographics plus hourly vitals and labs; data were split 80/10/10 at the patient level with forward-fill (≤6 h) then patient-specific median imputation (global median fallback for remaining missingness).

• Segmented each 24-hour trajectory into six 4-hour windows and constructed a per-window graph where nodes are features and adjacency matrices are learned from data (no predefined ontology); paired these graphs with LSTM per-feature embeddings and a pseudo-attention interpretability matrix.

• Trained end-to-end with a multi-loss objective combining contrastive graph augmentation, focal loss for class imbalance, and structural/regularisation terms; used a VGAE+GIN encoder, temporal pooling, early stopping, and five random seeds for robustness.

• Designed with deployment in mind: top-k sparsification and temporal pooling to limit compute, uses standard EHR feeds, produces time-specific importance scores, and is architected for millisecond inference on GPU/sidecar infrastructure (inference measured ≈13 ms in experiments).

• DynaGraph outperformed 14 baselines across 40,856 admissions. On MIMIC-III it achieved AUROC 0.856, AUPRC 0.461 and sensitivity 85.22% (+12 percentage points vs MedGNN); on eICU AUROC 0.802 (sensitivity 86.00%) and on HiRID AUROC 0.881 (sensitivity 86.20%).

• Relative AUPRC improved ≈6–8% versus state-of-the-art models (macro-averaged), but F1 scores remained modest (≈45–59%) and rare labels stayed challenging (for example, 31% precision at 50% recall for 30-day readmission, prevalence ≈2.2%).

• The pseudo-attention interpretability produced time-resolved importance patterns (early demographics → mid renal/electrolytes → late inflammatory/hemodynamic markers). The model learned strong edge couplings (e.g., creatinine–urea >0.8) and identified a creatinine–haemoglobin edge that increased ≈8–10 hours before AKI criteria — an association that suggests an earlier signal but does not establish causality.

• Practical performance notes: millisecond inference (≈13 ms); ablations showed removing contrastive augmentation or focal loss reduced AUROC by ≈0.048 and ≈0.035 respectively. The model maintained higher performance than MedGNN across the tested out-of-distribution subgroups (balanced accuracy), though subgroup sizes and clinical differences may limit generalisability conclusions.

• Treat DynaGraph as a high-sensitivity screener to prioritise pharmacist review, not as an autonomous decision engine—clinical judgement remains central because precision for rare events is modest.

• For pharmacists: interpret outputs as a 'physiology map' rather than just a score—the model surfaces time-specific drivers and feature pairings (e.g., creatinine–urea) that indicate which organ interactions are associated with current risk.

• Verify local fit: this evaluation used demographics, vitals, and labs but did not include medication-administration nodes. Before deployment, ask about lead-time versus current alerts, subgroup robustness on your units, and how explanations will be presented in workflows.

• Key operational checks before piloting: ensure sustained recall with manageable alert volume, explanations that align with clinical reasoning, measurable lead-time benefit in local data, integration of medication data, and light-touch drift monitoring with MLOps governance.

Strengths:

• Large multi-dataset evaluation with patient-level splits, out-of-distribution tests, five random seeds, and systematic ablations (datasets: MIMIC-III, eICU, HiRID, EHRSHOT; reported total 40,856 admissions).; Deployment-oriented engineering choices (top-k sparsification, temporal pooling), measured millisecond inference, and time-resolved pseudo-attention that support interpretability and potential operational use (subject to infrastructure and governance).

Limitations:

• No medication-administration nodes and a fixed 24-hour, six-window framing limit detection of drug-driven signals or longer-horizon patterns relevant to pharmacy interventions; as presented, the model functions primarily as a physiological monitor rather than a pharmacovigilance tool.; Retrospective-only evaluation, computationally intensive training, and demographic under-representation in source datasets (e.g., ≈8% Black in MIMIC-III) raise generalisability, fairness, and temporal-drift concerns that mandate local validation, subgroup evaluation, and MLOps monitoring before clinical deployment.

Treat DynaGraph as a high-sensitivity, time-resolved screening tool to prioritise pharmacist review; pursue local validation and pilots that emphasise alert-volume control, measurable lead-time benefit, inclusion of medication data, and deployment governance before using it to trigger automated interventions.