Graduation Date

Summer 8-14-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Programs

Biostatistics

First Advisor

Christopher Wichman Ph.D.

Second Advisor

Lynette Smith, Ph.D.

Third Advisor

Jianghu (James) Dong, Ph.D.

Abstract

Lung cancer remains a leading cause of cancer-related mortality worldwide. Accurate prognostic models are essential for identifying high-risk patients and informing treatment strategies. The present study replicated and extended the prognostic analysis of a 15-gene expression signature for early-stage non-small cell lung cancer (NSCLC). Publicly available datasets from Der et al. (2014) and Zhu et al. (2010) were utilized. The original principal component analysis (PCA)-Cox model was first replicated and validated. Subsequently, Ridge Cox, LASSO Cox, Elastic Net Cox, CoxBoost, Survival Support Vector Machine (Survival SVM), and Supervised PCA-Cox models were developed to compare their prognostic performance.

Patients were stratified into high-risk and low-risk groups based on model-derived risk scores. Kaplan–Meier survival curves, log-rank tests, and multivariable Cox proportional hazards regression were employed to evaluate survival differences between risk groups. Predictive performance was assessed using the concordance index (C-index) and the time-dependent area under the receiver operating characteristic curve (AUC).

The replication analysis reproduced the principal findings of the original study. Among penalized Cox regression models, Ridge Cox exhibited the highest predictive performance (C-index = 0.649, AUC = 0.712), followed by LASSO Cox (C-index = 0.635, AUC = 0.701) and Elastic Net Cox (C-index = 0.634, AUC = 0.699). Among machine learning approaches, CoxBoost achieved the highest predictive performance (C-index = 0.656, AUC = 0.724), followed by Supervised PCA-Cox (C-index = 0.622, AUC = 0.680), whereas Survival SVM demonstrated the lowest predictive performance (C-index = 0.538, AUC = 0.584). Overall, CoxBoost outperformed all other evaluated models.

In conclusion, this study suggests that the 15-gene expression pattern has value for predicting outcomes in early-stage NSCLC and that both penalized Cox regression and machine learning methods can successfully group patients by survival risk. Among all models tested, CoxBoost yielded the best predictive performance, suggesting that boosting-based survival models may improve outcome prediction using gene expression data.

Rights

The author holds the copyright to this work and any reuse or permissions must be obtained from the author directly.

Download

Available for download on Saturday, July 29, 2028

Share

COinS