IDDF2024-ABS-0431 Machine learning models in predicting survival for colorectal cancer patients with type 2 diabetes: a 20-year follow-up of 10,749 subjects

• Published in: Gut Vol. 73; no. Suppl 2; pp. A379 - A380
• Main Authors: Huang, Junjie, Jiang, Yu, Li, Yu, Dou, Qi, Huang, Ziwei, Zhong, Claire Chenwen, Hang, Junjie, Yuan, Jinqiu, Wong, Martin CS
• Format: Journal Article
• Language: English
• Published: London BMJ Publishing Group Ltd and British Society of Gastroenterology 01.08.2024; BMJ Publishing Group LTD
• Subjects: Cholesterol; Clinical Gastroenterology; Colorectal cancer; Colorectal carcinoma; Diabetes; Diabetes mellitus (non-insulin dependent); Diagnosis; Learning algorithms; Low density lipoprotein; Machine learning; Medical prognosis; Patients; Population studies; Risk factors; Risk groups; Survival; Survival analysis
• ISSN: 0017-5749; 1468-3288
• DOI: 10.1136/gutjnl-2024-IDDF.339

BackgroundPatients with type 2 diabetes (T2D) have an increased risk of developing colorectal cancer (CRC) compared to the general population. Several studies have suggested that the co-occurrence of these two chronic conditions can negatively impact patient outcomes, leading to poorer survival rates. This study aimed to identify factors that contribute to survival outcomes in this high-risk patient population to inform the development of more targeted and personalized management strategies to improve prognosis for CRC patients with pre-existing T2D.MethodsWe analyzed data from 10,749 participants in the Hong Kong Hospital Authority Data Collaboration Laboratory (HADLC) from 2000-2020. Various demographic and clinical variables were extracted, and several machine learning models, including Cox Proportional-Hazards (CoxPH), Lasso-regularized Cox (Coxnet), Random Survival Forest (RSF), Survival Decision Tree, and Gradient Boosting Survival Analysis (Coxboost), were used to predict CRC survival in T2D patients. Model performance was evaluated using the concordance index (C-index) and time-dependent area under the ROC curve (AUC).ResultsThe key predictors of CRC survival in T2D patients were T2D diagnosis age (HR=1.072, p<0.001), duration of T2D (HR=1.065, p<0.001), use of anti-lipid drugs (HR=0.796, p<0.001), central obesity (HR=0.830, p<0.001), and LDL cholesterol levels (HR=1.422, p<0.001). The Coxboost model achieved the highest C-index of 0.754 in the testing phase, outperforming the traditional CoxPH model (C-index=0.737). Shapley Additive Explanations (SurvSHAP) analysis showed that the duration of T2D, T2D diagnosis age, anti-lipid drug use, LDL cholesterol, and sex were the most influential factors on patient survival.ConclusionsThe machine learning models, particularly Coxboost and RSF, demonstrated superior performance in predicting CRC survival in T2D patients compared to the traditional CoxPH model. The identified risk factors can inform more targeted clinical management approaches for this high-risk patient population. By tailoring interventions based on these critical predictors, healthcare providers may be able to improve CRC survival rates among patients with concurrent T2D.