A Single-Center Study: XGBoost-Based Multi-Omics Prediction Model for Sepsis-Associated Acute Kidney Injury (SA-AKI) and Personalized Immune Therapy Efficacy in Pancreatic Cancer ICU Patients.

BACKGROUND

Sepsis-associated acute kidney injury (SA-AKI) is a highly prevalent and lethal complication in pancreatic cancer patients admitted to the intensive care unit (ICU), characterized by complex pathophysiology involving tumor-induced inflammation and immune dysfunction. Traditional prediction models lack specificity for this population, and targeted therapeutic strategies remain limited.Future multi-center prospective studies will further validate the model and improve its generalizability for clinical application in diverse patient populations.

METHODS

We conducted a single-center mixed retrospective-prospective cohort study involving 523 pancreatic cancer ICU patients (366 retrospective, 157 prospective) between January 2020 and December 2024. Clinical data, laboratory indices, natural killer (NK) cell activity, T-cell receptor (TCR) sequencing, and NK cell single-cell RNA sequencing (scRNA-seq) data were integrated. The XGBoost algorithm was used to construct a SA-AKI prediction model,with performance evaluated via area under the receiver operating characteristic curve (AUC),precision-recall curve (AUPRC), calibration curve, decision curve analysis (DCA), and Brierscore. SHAP (SHapley Additive exPlanations) analysis identified key biomarkers. Risk stratification (low/medium/high-risk) guided personalized immune therapy (NK cell infusion ± TCR-T therapy). Primary endpoints included 28-day ICU discharge rate, 90-day mortality, and overall survival (OS); secondary endpoints included biomarker dynamic changes and treatmentresponse.

RESULTS

The XGBoost model achieved an AUC of 0.959 (95% CI: 0.938-0.979) and AUPRC of 0.946 in the training cohort, with consistent performance in internal validation (AUC=0.923, AUPRC=0.905). Key predictive biomarkers included APACHE II score (mean absolute SHAP value=0.606), TCR clonality (0.504), NK cell activity (0.425), IFN-γ expression in NK cells (0.398), and CRP (0.161). Risk stratification revealed significant differences in responserates (82.1% vs. 52.4% vs. 34.2%), 28-day ICU discharge rates (85.9% vs. 69.5% vs. 41.6%),and 90-day mortality (25.9% vs. 47.4% vs. 71.2%) across low/medium/high-risk subgroups (all p<0.001). Personalized immune therapy significantly improved OS in medium/high-risk patients (median OS: 172/98 days vs. 115/65 days in controls, p<0.001). Dynamic monitoringshowed that NK cell activity (increase by 39%) and TCR clonality (increase by 25%) were significantly elevated post-treatment in responders (p<0.001). NK cell scRNA-seq identified functional (IFN-γ+GZMB+), exhausted (PD-1+), and naive subgroups, with higher functional NK cell proportion in responders (36.0% vs. 13.9%, p<0.001).

CONCLUSIONS

The XGBoost-based multi-omics model enables accurate SA-AKI prediction and risk stratification in pancreatic cancer ICU patients. Risk-stratified personalized immune therapy improves clinical outcomes, and dynamic monitoring of NK cell activity/TCR clonalityserves as a surrogate for therapeutic efficacy. This study provides a novel clinical decision-making pathway for SA-AKI management in this high-risk single-center population.