Background and Aims:
Colon cancer remains a leading cause of cancer-related mortality globally. This study aimed to identify
prognostic factors and evaluate predictive models for survival in colon cancer patients treated at Imam Khomeini Hospital, Sari, Iran. Methods: A retrospective cohort analysis was conducted on ۱,۱۰۶ patients diagnosed with colon cancer between ۲۰۱۲–۲۰۱۷, with follow-up until April ۲۰۲۱. Data were collected via a validated checklist, including demographic, clinical, and treatment variables. Survival analysis employed Kaplan-Meier curves, Cox proportional hazards regression, and
machine learning models (SVM, Random Survival Forest,
Gradient Boost Survival). Model performance was assessed using the concordance index (C-index) and time-dependent AUC. Results: Among ۱,۱۰۶ patients, ۲۳۴ (۲۱.۲%) experienced mortality. The KM curve revealed a progressive decline in survival rates, from ۸۰% at ۵ months to ۲۰% by ۵۰ months, highlighting the critical impact of early intervention. Also, Kaplan-Meier analysis stratified by survival quality revealed markedly lower survival in patients with impaired activity versus normal activity.
Cox regression identified significant predictors of survival: synchronous CRC cancer (HR=۰.۱۷, ۹۵% CI:۰.۰۵–۰.۵۵, p<۰.۰۰۵), IV drug use (HR=۲۱.۶۲, ۹۵% CI:۲.۸۶–۱۶۳.۴۹, p<۰.۰۰۵), blood per rectum (HR=۱.۶۷, ۹۵% CI:۱.۲۰–۲.۳۱, p<۰.۰۰۵), marital status (married vs. unmarried: HR=۰.۵۳, ۹۵% CI:۰.۳۵–۰.۷۹, p<۰.۰۰۵), inflammatory bowel disease (HR=۴.۹۳, ۹۵% CI:۱.۸۳–۱۳.۲۸, p<۰.۰۰۵), and survival quality (normal activity vs. impaired: HR=۰.۱۰, ۹۵% CI:۰.۰۷–۰.۱۴, p<۰.۰۰۵). Machine learning models demonstrated strong predictive performance, with the
Random Survival Forest achieving the highest C-index (۰.۷۹۱), followed by
Gradient Boost (۰.۷۴۵),
SVM (۰.۷۴۴), and Cox PH (۰.۷۲۳). Time-dependent AUC curves indicated robust discrimination across follow-up periods. Conclusion: Key prognostic factors, including comorbidities, lifestyle factors, and treatment response, significantly influence colon cancer survival. Machine learning models, particularly Random Survival Forest, outperformed traditional
Cox regression in prediction accuracy. These findings underscore the clinical utility of integrating patient-specific risk factors and advanced predictive models to optimize survival stratification and personalized care. Further validation in multi-center cohorts is warranted to enhance generalizability. The KM plot underscores temporal mortality trends, reinforcing the need for timely clinical management. These insights advocate
