Development of machine learning predictive models and risk scoring system for survival in breast cancer patients with type II diabetes: a retrospective cohort study

<h2>Abstract</h2><h3>Background</h3> Breast cancer patients with type II diabetes mellitus (T2DM) encounter distinct challenges that can influence their survival outcomes. Developing machine learning predictive models and a comprehensive risk scoring system is crucial for identifying associated risk factors and enhancing survival probabilities in this population. This study aims to utilize machine learning techniques to develop predictive models and create a risk score system to understand the risk factors associated with survival in breast cancer patients with T2DM and to estimate their survival probabilities. <h3>Methods</h3> Data were obtained from the Hong Kong Hospital Authority Data Collaboration Laboratory, covering 2000 to 2020. Five machine learning algorithms were employed to develop predictive models for the survival of breast cancer patients with T2DM: Cox proportional hazards regression, LASSO Cox regression, boosting, survival tree, and random survival forest (RSF). Model performance was evaluated using a time-dependent area under the curve (AUC) and concordance index (C-index). The best-performing model identified key risk factors through SHAP (Shapley Additive Explanations) analysis, while the AutoScore-Survival package facilitated the creation of a risk score system. <h3>Findings</h3> This retrospective cohort study included 8,255 breast cancer patients with T2DM. The mean survival time was 97.83 months (SD: 70.39), with 99.21% female and 16.00% deceased. The RSF model demonstrated the best predictive performance, achieving an AUC of 0.840 and a C-index of 0.79. The risk score system was established based on following criteria: age at breast cancer diagnosis (0 points for <50 years, 12 for 50-60, 21 for 60-70, and 40 for ≥70), T2DM duration (0 points for <1 year, 20 for 1-5 years, and 24 for ≥5 years), high-density lipoprotein cholesterol (HDL_C), (3 points for <1 mmol/L and 0 for ≥1 mmol/L), low-density lipoprotein-cholesterol (LDL-C) (0 points for <3.4 mmol/L and 8 for ≥3.4 mmol/L), and creatinine levels (18 points for <47 μmol/L, 1 for 47-56, 0 for 56-81, 5 for 81-115, and 10 for ≥115 μmol/L). This system identified 48.6% of patients as high-risk with scores >65, corresponding to a 5-year survival probability of 27.7%. <h3>Interpretation</h3> These findings underscore the importance of the risk score system, particularly the roles of HDL-C, LDL-C, and creatinine levels as critical factors in identifying high-risk breast cancer patients with T2DM. By integrating these biomarkers into clinical assessments, healthcare providers can better tailor interventions and management strategies to improve survival outcomes. Future research should focus on validating the risk score across diverse populations and exploring targeted therapies for patients identified as high-risk based on these key indicators.