A nomogram-based radiomics for predicting survival to concurrent chemoradiotherapy in inoperable pancreatic cancer: a dual-center cohort study

Xin Liu¹Ke Su²Shanshan Du¹Peiping Sun¹Shucheng Shen¹Benzhe Liang¹Jian Chen¹Rui Liu¹Rui Zhang³Heran Wang⁴Huadong Wang¹Yong Yin¹Zhenjiang Li^1*

• ¹Department of Radiation Oncology, Shandong Cancer Hospital, Jinan, China
• ²Southwest Medical University, Luzhou, China
• ³Nanjing University Medical School Affiliated Nanjing Drum Tower Hospital, Nanjing, China
• ⁴Shengjing Hospital of China Medical University, Shenyang, China

Objective: This study was designed to explore the value of machine learning-based radiology in predicting overall survival (OS) among patients with inoperable pancreatic cancer (PC) who are undergoing concurrent chemoradiotherapy (CCRT). Methods: This multicenter study enrolled 342 patients with inoperable PC. Firstly, radiomic features were pre-screened by univariate Cox regression and subsequently used to develop 101 machine-learning–based imaging models. An optimized selection algorithm was applied to these models to derive each patient's radiomic signature (Rad-score). Secondly, key clinical predictors of OS were identified via LASSO–Cox regression and incorporated into clinical nomogram. Finally, the Rad-score was combined with the independent clinical risk factors to construct clinical–radiomics nomogram. Results: LASSO–Cox regression identified age, clinical stage, tumor size, and albumin level as independent prognostic factors for OS. Based on these four variables, we constructed a clinical nomogram in the training cohort, which achieved a C-index of 0.71. In the internal validation cohort, the areas under the receiver operating characteristic curve (AUC-ROC) for predicting 1-, 2-, and 3-year OS were 0.577, 0.721, and 0.730, respectively; in the external validation cohort, the corresponding AUC-ROCs were 0.841, 0.757, and 0.598. Subsequently, each patient's Rad-score was integrated with these clinical predictors to develop a clinical–radiomics nomogram, which demonstrated a C-index of 0.892. The AUC-ROCs for predicting 1-, 2-, and 3-year OS were 0.791, 0.846, and 0.840 in the internal validation cohort, and 0.863, 0.830, and 0.734 in the external validation cohort. Conclusion: The clinical–radiomics nomogram demonstrated superior predictive performance for OS compared to the clinical nomogram in inoperable PC patients undergoing CCRT.