TL;DR
This study developed machine learning models to predict survival in uterine cancer patients with Type 2 Diabetes (T2D), identifying key risk factors and creating a practical risk scoring system. The findings highlight the potential of AI to personalize treatment and improve outcomes for this vulnerable patient group.
What Problem Does This Paper Address?
Uterine cancer is a significant global health concern, and its prognosis is further complicated by the increasing prevalence of Type 2 Diabetes (T2D). While T2D is known to increase the risk of all-cause and cancer-specific mortality in uterine cancer survivors, the specific interplay of risk factors and their impact on survival in this combined patient population remains poorly understood. Existing prognostic models often struggle with the complexity of high-dimensional data and diverse risk factors. This paper aims to bridge this gap by leveraging machine learning (ML) to develop accurate predictive models and a clinically actionable risk scoring system for uterine cancer patients with T2D, enabling more personalized clinical management.
Methods
The researchers employed a retrospective, territory-wide cohort study design, utilizing data from the Hong Kong Hospital Authority Data Collaboration Laboratory (HADCL). To develop predictive models, they explored several machine learning algorithms, including:
- Cox Proportional Hazards (CoxPH) Regression: A standard statistical method for survival analysis.
- LASSO Cox Regression (CoxNet): A penalized regression technique for feature selection and regularization.
- Survival Tree (ST): A tree-based model for survival prediction.
- Boosting: An ensemble method that combines multiple weak learners.
- Random Survival Forest (RSF): An ensemble of decision trees specifically designed for survival data, known for its ability to capture complex, non-linear relationships.
Key risk factors were identified using Shapley Additive Explanations (SHAP) analysis, a method that provides a unified approach to explaining the output of any machine learning model. For creating the risk scoring system, they utilized the AutoScore-Survival package, an extension of the AutoScore framework designed for time-to-event data. This package integrates ML with Cox regression to generate parsimonious and interpretable models.
Datasets
The study utilized data from the Hong Kong Hospital Authority Data Collaboration Laboratory (HADCL), specifically drawing from the Risk Assessment and Management Programme for Diabetes Mellitus (RAMP-DM). This dataset included information on over 2000 adult patients diagnosed with both uterine cancer and T2D between 2000 and 2020. The data was split into training (70%) and testing (30%) sets for model development and validation. No new benchmarks were introduced; the focus was on applying existing ML techniques to a specific clinical cohort.
Key Insights and Ideas
- RSF Dominance: The Random Survival Forest (RSF) model emerged as the top performer, significantly outperforming other ML algorithms in predictive accuracy. This suggests that complex, non-linear interactions between risk factors are crucial for predicting survival in this population.
- Key Predictors Identified: The study pinpointed several critical factors influencing survival, including:
- Age at cancer diagnosis: Older age at diagnosis is associated with increased mortality risk.
- Duration of T2D: A longer history of diabetes exacerbates the risk.
- BMI and Creatinine Levels: Elevated levels of these markers are linked to poorer survival.
- Central Obesity: Interestingly, central obesity was found to be negatively associated with mortality risk, suggesting a potentially protective effect in this specific context, which warrants further investigation.
- Interpretable Risk Score: The development of the AutoScore-Survival risk scoring system is a significant practical contribution. It translates complex ML findings into a simple, score-based system that clinicians can readily use to stratify patients into high- and low-risk groups.
- Clinical Utility: The risk score system effectively stratified patients, with a high-risk group (scoring >40) exhibiting a substantially lower 5-year survival probability (43.5%) compared to the low-risk group. This demonstrates its potential for guiding individualized treatment and management strategies.
Key Results
- The study included 2047 uterine cancer patients with T2D.
- The RSF model achieved the highest predictive performance, with a time-dependent AUC of 0.823 and a C-index of 0.90.
- The developed risk scoring system identified 31.4% of patients as high-risk, who had a 5-year survival probability of 43.5%, approximately 1.7 times lower than the low-risk group.
- Key predictors identified for increased mortality risk included age at cancer diagnosis, duration of T2D, BMI, and creatinine levels. Central obesity was associated with a decreased risk of mortality.
- The risk score was based on seven factors: age at cancer diagnosis, duration of T2D, creatinine levels, serum potassium level, LDL-C level, BMI, and triglycerides level.
Limitations and Future Work
The authors acknowledge several limitations:
- Missing Data: The study did not include crucial variables like tumor subtype, FIGO stage, or treatment adherence, which could further refine the models.
- Generalizability: The reliance on data from the Hong Kong public healthcare system might limit the applicability of the findings to other healthcare contexts.
- External Validation: While internal validation was performed, external validation across diverse populations was not conducted.
- Retrospective Design: The retrospective nature of the study limits causal inference.
Future work suggested includes:
- Incorporating more clinical data: Including tumor characteristics and treatment details to enhance the risk score system.
- Developing an open-access web calculator: To facilitate clinical implementation and broader testing of model generalizability.
- Investigating underlying mechanisms: Further research is needed to understand the biological pathways driving the observed associations.
- External validation: To confirm the robustness and applicability of the risk score system across different populations.
Summary updated