Prediction models of locally advanced rectal cancer prognosis incorporating perioperative longitudinal clinical data: A retrospective longitudinal cohort study.

Abstract

e15624 Background: For locally advanced rectal cancer, neoadjuvant chemoradiation(CCRT) followed by surgery is a standard treatment. Clinical T/N stage, whether pathological complete remission (pCR) after neoadjuvant CCRT or not are well recognized tumor features that predict disease prognosis. Although the parameters that determine pathologic stage are the strongest predictors of postoperative outcome, other clinical, molecular, and histologic features may influence prognosis independent of stage. The aim is to compare the predictive efficiency, between different models for overall survival in rectal cancer patients. Methods: This retrospective study is approved by TMU-JIRB(Protocol Number: N201809032).This is also a single institute study. 175 newly-diagnosed patients with locally advanced rectal cancer were retrospectively enrolled between Jan.2011 and Dec.2016. Basic information include sex, age, body weight, body height, previous medical history were recorded. CEA level, chemotherapy regimen, type of surgery, and distance between tumor and anal verge were also collected. Chemotherapy regimen of CCRT including 5FU+LV, capecitabine (Xeloda), FOLFOX, and FOLFIRI are recorded. Surgical resection was then followed 4-8 weeks after completion of CCRT. The 2010 AJCC tumor response grading (TRG) system was used to report treatment response of proctectomy samples. The progression free survival (PFS) and overall survival (OS) are then calculated. PFS is defined as survival without any recurrence or distant metastasis before Dec 31st, 2023. A comprehensive evaluation was conducted by incorporating all these machine learning models together. The aim was to compare their predictive efficiency, with models such as Random Forest, Naïve Bayes, KNeighbor, SVM, Decision Tree, Logistic Regression, XGBoost, and LightGBM. Results: 175 patients (104 men and 71 women) were enrolled in this study. The median follow up time was 7.1 yrs. The BMI was 24.05 ± 0.03 for men and 24.20 ± 0.07 for women. 32 (17.3%) patients achieved pCR, TRG 0. Clinical staging cT3 (78.8%) and nodal positive (78.8%) are the most common tumor feature in this study group. The median age of male and female patient is 58.6 and 56.1 respectively. Upon applying various machine learning models, different predictive accuracies emerged. The Random Forest, KNeighbor, and SVM models all achieved an accuracy of 0.9250. Naïve Bayes had an accuracy of 0.8620, while the SVC model registered at 0.9310, and XGBoost was slightly behind at 0.9286. SVC model is marking the best-performing model in this evaluation. Conclusions: This multicenter retrospective data-driven deep learning prediction model informed that SVC model yielded the best predictive results. Further prospective clinical trials and external validation of this prediction model are needed to further reinforce our findings.