Building Machine Learning Models Based on Oxidative Stress Index Score to Predict Survival of Locally Advanced Rectal Cancer Receiving Different Neoadjuvant Therapy Patterns.

BACKGROUND: Liver enzyme biomarkers are known to contribute to the onset and progression of colorectal cancer. OBJECTIVE: To develop a novel oxidative stress index score integrating γ-glutamyl transferase and total bilirubin, evaluate its prognostic value in locally advanced rectal cancer patients receiving neoadjuvant therapy, and construct and validate machine learning-based survival prediction models incorporating oxidative stress index score. DESIGN: A novel liver enzyme indicator - oxidative stress index score was established by integrating γ-glutamyl transferase and total bilirubin. Machine learning models were constructed based on oxidative stress index score and clinicopathological characteristics to predict survival. The predictive performance of these models was evaluated and interpreted and further validated in the validation cohort. Using these models, patients were stratified into low risk and high-risk groups for both disease-free survival and overall survival, respectively. SETTINGS: Data were collected from Sun Yat-Sen University Cancer Center between May 2007 and August 2018. PATIENTS: We enrolled locally advanced rectal cancer patients who had undergone either total neoadjuvant therapy (as the training cohort) or neoadjuvant chemoradiotherapy (as the validation cohort), followed by total mesorectal excision with or without adjuvant chemotherapy. MAIN OUTCOME MEASURES: Disease-free survival and overall survival in the training and validation cohort. RESULTS: A total of 970 locally advanced rectal cancer patients were allocated to a training cohort (n = 521) and a validation cohort (n = 449). Among 5 machine learning models, random survival forests demonstrated optimal predictive efficiency. Thus, the random survival forests model was constructed based on oxidative stress index score. The random survival forests models outperformed ypstage in prediction efficiency significantly across both cohorts. In both cohorts, low-risk groups exhibited significantly higher disease-free survival and overall survival rates compared to high-risk groups (all p < 0.000 1). A free online random survival forests calculator has been developed (https://zhuoliworkroom.shinyapps.io/new_shiny/). LIMITATIONS: Our study has inherent limitations of retrospective and single-center studies. CONCLUSIONS: A higher oxidative stress index score was significantly associated with worse prognostic outcomes. Collectively, these random survival forests models exhibit superior predictive performance compared to conventional ypstage indicators. See Video Abstract.