Abstract
Traditional Cox proportional-hazards models are the standard for identifying prognostic factors in oncology, yet they often fail to account for "competing risks"—events like death from non-cancer causes that prevent the primary event of interest. In thyroid carcinoma, where survival is often prolonged, these competing risks are significant. This study analyzed 38,444 patients from the SEER database using three multivariate models: Cox regression, cause-specific hazard (CS), and subdistribution hazard (SD). Results identified age, histology, stage, and metastasis as consistent risk factors. Notably, the competing-risks models revealed that traditional Cox models might overestimate the impact of race while underestimating the significance of tumor size. These findings suggest that competing-risks analysis provides a more accurate framework for clinical risk stratification in thyroid cancer.
Introduction
Predicting survival outcomes for thyroid carcinoma is complex due to the disease's generally favorable prognosis and long clinical course. In such scenarios, patients are statistically likely to die from causes unrelated to their cancer (e.g., cardiovascular disease) before the cancer itself becomes terminal. Standard statistical tools, such as the Cox proportional-hazards model, treat these "competing" deaths as censored data, which can lead to biased results and an overestimation of cancer-specific mortality. This study aims to refine prognostic accuracy by applying competing-risks models to a large-scale patient cohort.
Methods
The researchers utilized data from the Surveillance, Epidemiology, and End Results (SEER) database, identifying 38,444 eligible patients diagnosed with thyroid carcinoma between 2010 and 2015. The cohort included various histological types: Papillary (PTC), Follicular (FTC), Medullary (MTC), and Anaplastic (ATC).
To identify prognostic factors, three statistical approaches were compared:
- Cox Regression Analysis: The standard model for survival data.
- Cause-Specific (CS) Hazard Function: Measures the "force" of a specific cause of death.
- Subdistribution (SD) Hazard Function: Estimates the actual probability of an event in a real-world setting where multiple risks coexist. This model is often preferred for clinical decision-making.
Results
The analysis highlighted several factors that consistently influenced prognosis across all three models.
Consistent Risk Factors
- Demographics and Social Factors: Increasing age was a primary risk factor. Conversely, being female and being married were identified as protective factors.
- Histology: Compared to Papillary carcinoma, Anaplastic (ATC) was the most aggressive, followed by Medullary (MTC) and Follicular (FTC).
- Clinical Staging: Higher AJCC-6 stages (II, III, and IV) and the presence of liver or lung metastases were strong predictors of mortality.
- Treatment and Staging Quality: Lack of surgery and radioiodine therapy were significant risks. Interestingly, having regional lymph nodes examined—regardless of the findings—was a protective factor, likely serving as a proxy for thorough surgical care.
Discrepancies in Traditional Modeling
The study found that competing-risks models corrected several biases inherent in the Cox model:
- Race: While the Cox model suggested that being Black was a significant risk factor (HR 1.25), this significance disappeared in the CS and SD models, suggesting the initial finding was an artifact of non-cancer-related mortality risks.
- Tumor Size: The Cox model failed to find statistical significance for tumor sizes >1 cm. However, both competing-risks models (CS and SD) identified larger tumor size as a highly significant risk factor (p < 0.01).
Discussion and Conclusion
By accounting for deaths due to other causes (DOC), which represented nearly half of the total mortality in this cohort, the researchers demonstrated that competing-risks models provide a more nuanced view of thyroid cancer prognosis. The subdistribution hazard (SD) model, in particular, offers a more conservative and likely more accurate estimation of risk.
These findings underscore the importance of using appropriate statistical methods when dealing with high-survival cancers. The identified factors—including the protective effects of marriage and thorough lymph node examination—provide a more robust evidence base for treatment planning and patient counseling. Future work should validate these results using the AJCC 8th edition staging and incorporate more granular treatment data, such as targeted molecular therapies.
Summary updated