.png){kind=icon}
.png){kind=icon}
Return
Peripheral endocrine–nutritional machine learning signature predicts short-term response to neoadjuvant immunochemotherapy in locally advanced gastric cancer
.png){kind=icon}
OA
AI
F
Z
H
T
P
Z
Y
W
Y
W
Q
DOI：10.3389/fimmu.2026.1907116.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
BackgroundMajor pathological response (MPR) to neoadjuvant PD-1 inhibitor plus chemotherapy in locally advanced gastric cancer (LAGC) varies widely and is incompletely explained by tumor-side biomarkers. Because checkpoint inhibitors act within a host environment shaped by endocrine; nutritional; and inflammatory state; we developed and validated the Endocrine–Nutritional Immunotherapy response score (ENI-Score) from routine pretreatment peripheral-blood markers to predict short-term response and early recurrence.MethodsWe retrospectively analyzed 738 patients with LAGC or gastroesophageal junction adenocarcinoma who received neoadjuvant PD-1 inhibitor plus platinum-based chemotherapy and radical gastrectomy at four hospitals; partitioned into training (n=335); internal (n=114); and external (n=289) validation cohorts. Ten machine-learning algorithms were benchmarked; and a prespecified; domain-balanced five-marker panel comprising 8 AM cortisol; prognostic nutritional index (PNI); prealbumin; neutrophil-to-lymphocyte ratio (NLR); and C-reactive protein-to-albumin ratio (CAR) was carried forward in a regularized logistic model and rescaled to the ENI-Score (0 -100). Analyses included SHAP; tertile stratification; nested AUC comparison (DeLong); calibration; decision-curve analysis; and Kaplan–Meier survival analysis; the primary endpoint was MPR (residual viable tumor ≤10%).ResultsThe overall MPR rate was 40.9% (302/738). All five markers differed significantly between MPR and non-MPR patients in a biologically coherent direction consistent across all three cohorts (each P<0.001). Regularized logistic regression achieved the highest; most stable external-validation AUC among the ten algorithms (training 0.873; internal 0.816; external 0.830; and five-fold cross-validated 0.862); whereas tree-based ensembles overfit (training AUC up to 0.99; external ≤0.81). ENI-Score tertiles separated MPR rates steeply (Low 10.4%; Intermediate 38.2%; High 79.1%; P-trend<0.001); with an adjusted High-benefit versus Low-benefit odds ratio of 35.3 (95% CI 20.6-60.5) and a per 10-point adjusted OR of 1.72 (95% CI 1.59-1.86; P<0.001). The ENI-Score reached an overall AUC of 0.842 and significantly augmented clinical staging (0.656 to 0.864; DeLong P<0.001). Higher ENI-Score was associated with longer recurrence-free survival (RFS; log-rank P<0.001; adjusted HR per 10 points 0.78).ConclusionsBuilt entirely from five low-cost routine peripheral-blood markers; the ENI-Score accurately and reproducibly stratifies MPR after neoadjuvant immunochemotherapy in LAGC; augments tumor-centered clinical staging; and is associated with RFS. It provides an inexpensive; interpretable host-state adjunct to tumor-side biomarkers that warrants prospective validation.
Keywords:
machine learning
gastric cancer
risk stratification
neoadjuvant immunochemotherapy
major pathological response
host endocrine-nutritional state
Journal
IF:
5.9
Papers:
4.9W
Citations:
22.7W
Organization
Cited Papers