Pan-cancer G2C-Pro: A two-stage Gaussian clustering to prognostically stratify patients with advanced tumors treated with immune checkpoint inhibitors

• Published in: Journal of clinical oncology Vol. 42; no. 16_suppl; p. e13605
• Main Authors: Paoloni, Francesco, Pecci, Federica, Bruschi, Giulia, Tola, Elisabetta, Sbrollini, Agnese, Galassi, Tommaso, Borgheresi, Alessandra, Cognigni, Valeria, Cantini, Luca, Santamaria, Luca, Gualtieri, Mariangela, Lunerti, Valentina, Agostinelli, Veronica, Agostini, Andrea, Mentrasti, Giulia, Giovagnoni, Andrea, Burattini, Laura, Berardi, Rossana
• Format: Journal Article
• Language: English
• Published: 01.06.2024
• ISSN: 0732-183X; 1527-7755
• DOI: 10.1200/JCO.2024.42.16_suppl.e13605

e13605 Background: Immune checkpoint inhibitors (ICIs) have revamped the clinical outcomes of patients (pts) affected by advanced solid tumors. The aim of our study is to develop a two-stage Gaussian Clustering algorithm (G2C) integrated with a logistic regression model (Pro) to prognostically stratifypts with advanced solid tumors treated with ICIs (pan-cancer G2C-Pro) based on baseline features. Methods: Data extraction involved, retrospectively, pts with advanced solid tumors (lung, melanoma, renal cancer, head and neck, urothelial carcinoma) treated with ICIs at Department of Medical Oncology, Ancona. Baseline body mass composition (BC) was assessed through computed tomography (CT) scan at L3 level, abstracting subcutaneous and visceral fat and muscle mass indicators. Moreover, baseline clinicopathologic features, nutritional status through The Controlling Nutritional Status (CONUT) score, and comorbidities were collected. An unsupervised clustering analysis was used to identify two groups of pts with different BC phenotype risk groups. Then, another unsupervised clustering analysis was used to identify two groups of pts within the dataset that hold prognostic significance according to BC risk groups, clinicopathologic features, CONUT score, and comorbidities. G2C was constructed by sequentially integrating two Gaussian Mixture models, each employing K-means++ initialization. Next, Pro was used to predict clusters’ label. The metrics used to evaluate the average performance of Pro on test sets were ACCuracy (ACC) and Area Under the Curve (AUC). The model was developed in Python on-cloud using the Google Colab service. Results: A total of 87 pts with complete data available were included in the final analysis. The two generated clusters for BC phenotype were BC_Low_Risk (n = 39) and BC_High_Risk (n = 48). Then, considering BC risk groups, clinicopathologic features, nutritional status, and comorbidities, two generated clusters were cluster 1 (n = 55) and cluster 2 (n = 32). Looking at clinical outcomes, median progression free survival was 16.1 months for cluster 1 versus 7.1 months for cluster 2 (HR: 0.57, 95% CI: 0.33-0.98, p = 0.04), and median overall survival was 41.8 months for cluster 1 versus 10.5 months for cluster 2 (HR: 0.47, 95% CI: 0.27-0.83, p = 0.008). The average ACC and AUC across all splits achieved by Pro model for patient classification into clusters were 0.94 and 0.89, respectively. From the feature ranking, it emerged that the one with higher importance was CONUT score, followed by BC phenotype risk groups and neutrophil-to-lymphocyte ratio. Conclusions: By using an easy-to-obtain and reproducible baseline BC, clinicopathologic, nutritional features, and comorbidities, we demonstrated that pan-cancer G2C-Pro is a promising and accurate prognostic model for stratifying pts with advanced tumors treated with ICIs.