PathSepsisNet: a pathway-aware temporal graph attention network integrating inflammatory crosstalk dynamics for early prediction of multi-organ dysfunction in sepsis.

Abstract

Sepsis-induced multi-organ dysfunction syndrome (MODS) remains a leading cause of intensive care unit (ICU) mortality worldwide. The complex interplay among inflammatory signaling pathways TLR4/NF-κB, NLRP3/P2X7 and Nrf2/HO-1 plays a critical role in determining organ damage trajectories, yet existing artificial intelligence (AI) models fail to capture these molecular dynamics. We propose PathSepsisNet, a pathway-aware temporal graph attention network that integrates biological pathway crosstalk knowledge into a deep learning architecture for early MODS prediction. We developed a biologically-informed graph neural network in which six nodes represent molecular pathway activation states (estimated from routine clinical biomarkers) and twelve directed edges encode established crosstalk relationships. A dual spatial-temporal attention mechanism with gated fusion captures inter-pathway crosstalk dynamics and within-pathway temporal evolution simultaneously. A novel Pathway Divergence Score (PDS) quantifies the imbalance between pro-inflammatory and cytoprotective pathways. The model was trained and validated on 29,765 adult patients with sepsis (Sepsis-3) from the Medical Information Mart for Intensive Care IV (MIMIC-IV) v3.1 using only the first 12 h of ICU data to predict MODS development at 12-36 h, with strict data leakage prevention (no SOFA sub-scores as inputs). PathSepsisNet achieved area under the receiver operating characteristic curve (AUROC) 0.895 and an area under the precision-recall curve (AUPRC) of 0.901 for MODS prediction, outperforming bidirectional long short term memory (BiLSTM) (0.886), Transformer (0.890) and logistic regression (0.876), while providing unique pathway-level interpretability. Gradient boosting methods (XGBoost: 0.906; LightGBM: 0.905) achieved marginally higher raw AUROC, but lack mechanistic insight. The PDS demonstrated remarkable standalone discriminative ability (MODS + : 0.596 ± 0.185 vs. MODS - : 0.299 ± 0.163; PDS AUROC: 0.872). Node importance analysis identified CytokineStorm (0.290), NLRP3/P2X7 (0.212) and mitochondrial dysfunction (0.211) as the most influential pathway modules. MODS + patients exhibited significantly higher lactate (2.34 vs. 1.63 mmol/L), CRP (123.5 vs. 97.2 mg/L), bilirubin (2.91 vs. 0.97 mg/dL) and lower albumin (2.82 vs. 3.04 g/dL) and platelets (160 vs. 207 K/μL), validating pathway-specific hypotheses. PathSepsisNet demonstrates that integrating biological pathway knowledge into graph neural network architectures yields predictive performance competitive with state-of-the-art methods, while uniquely providing pathway-level mechanistic insights. The PDS represents a novel computational biomarker bridging molecular pathway biology and clinical decision support. These findings support the paradigm that biologically informed AI can close the gap between predictive accuracy and clinical interpretability in sepsis management.