A Novel Machine Learning Algorithm Combined With Multivariate Analysis for the Prognosis of Renal Collecting Duct Carcinoma

• Published in: Frontiers in oncology Vol. 11; p. 777735
• Main Authors: Wei, Liwei, Huang, Yongdi, Chen, Zheng, Li, Jinhua, Huang, Guangyi, Qin, Xiaoping, Cui, Lihong, Zhuo, Yumin
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Media S.A 13.01.2022
• Subjects: collecting duct carcinoma; machine learning; Oncology; predictive model; renal cell carcinoma; SEER database; SEER database; renal cell carcinoma; predictive model; machine learning; collecting duct carcinoma
• ISSN: 2234-943X; 2234-943X
• DOI: 10.3389/fonc.2021.777735

To investigate the clinical and non-clinical characteristics that may affect the prognosis of patients with renal collecting duct carcinoma (CDC) and to develop an accurate prognostic model for this disease. The characteristics of 215 CDC patients were obtained from the U.S. National Cancer Institute's surveillance, epidemiology and end results database from 2004 to 2016. Univariate Cox proportional hazard model and Kaplan-Meier analysis were used to compare the impact of different factors on overall survival (OS). 10 variables were included to establish a machine learning (ML) model. Model performance was evaluated by the receiver operating characteristic curves (ROC) and calibration plots for predictive accuracy and decision curve analysis (DCA) were obtained to estimate its clinical benefits. The median follow-up and survival time was 16 months during which 164 (76.3%) patients died. 4.2, 32.1, 50.7 and 13.0% of patients were histological grade I, II, III, and IV, respectively. At diagnosis up to 61.9% of patients presented with a pT3 stage or higher tumor, and 36.7% of CDC patients had metastatic disease. 10 most clinical and non-clinical factors including M stage, tumor size, T stage, histological grade, N stage, radiotherapy, chemotherapy, age at diagnosis, surgery and the geographical region where the care delivered was either purchased or referred and these were allocated 95, 82, 78, 72, 49, 38, 36, 35, 28 and 21 points, respectively. The points were calculated by the XGBoost according to their importance. The XGBoost models showed the best predictive performance compared with other algorithms. DCA showed our models could be used to support clinical decisions in 1-3-year OS models. Our ML models had the highest predictive accuracy and net benefits, which may potentially help clinicians to make clinical decisions and follow-up strategies for patients with CDC. Larger studies are needed to better understand this aggressive tumor.