A machine learning approach to risk-stratification of gastric cancer based on tumour-infiltrating immune cell profiles

Résumé fourni par la source

Background Gastric cancer (GC) is a highly heterogeneous disease, and the response of patients to clinical treatment varies substantially. There is no satisfactory strategy for predicting curative effects to date. We aimed to explore a new method for predicting the clinical efficacy of GC treatment based on immune variables detected via flow cytometry.Methods We collected 394 tumour tissues from GC patients for flow cytometry analysis and gating analysis of tumour-infiltrating immune cells (TIICs). Unsupervised consensus clusters were generated from the cohort to classify patients into different phenogroups, and their clinical characteristics were examined. The derived model was evaluated via principal component analysis and t-distributed stochastic neighbourhood embedding analysis. Kaplan–Meier’s curves were used to determine the prognosis during a 920-day-long median follow-up period (interquartile range: 834–1071 days). Adjusted multivariate Cox regression analysis was used to evaluate the association of clusters with disease-free survival (DFS) and recurrence.Results All patients were classified based on their TIIC profiles into the C1 (characterized by low CD45 negative cell, high lymphocyte, high neutrophil and low CD3 + T cell levels), C2 (characterized by high CD8 + CD279+ cell and low CD4+ Th and CD8+ Tc cell numbers) and C3 (characterized by high CD4 + CD25+ and Treg cell levels) phenogroups. Patients from the three clusters had varied pathologies, MMR statuses and TIIC distribution patterns (p < .05). Kaplan–Meier’s analysis showed that the prognosis of C3 was inferior compared to C1 and C2 (p = .0025). Adjusted Cox proportional hazard models helped us identify that C1 and C2 exhibited a favourable factor of recurrence after surgery, compared to C3. Kaplan–Meier’s analysis showed that C1 and C2 were associated with a better DFS than C3 in some GC patient subgroups.Conclusions The machine learning model developed was found to be effective model at predicting the prognosis of patients with GC and their TIIC profiles for risk stratification in clinical settings.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.