INTRODUCTION:

Cancer remains a leading cause of morbidity and mortality worldwide, with over 10 million deaths recorded annually. While substantial progress has been made in cancer diagnosis and treatment, the burden of early mortality—deaths occurring within the first months following diagnosis—continues to represent a major public health concern, especially in high-burden cancers such as Thoracic(Lung and Bronchus), Upper Digestive(Pancreas, Liver), Lower Digestive(Rectum, Ascending Colon, Sigmoid Colon) and breast.

Problematic: Traditional survival metrics often emphasize 5-year survival rates, which can overlook critical short-term outcomes. Early mortality is not only a marker of aggressive disease but also a reflection of delayed diagnosis, barriers to timely care, and disparities in treatment access. These deaths, many of which may be preventable, highlight the need to better understand the determinants and thresholds that influence survival in the first few months after cancer diagnosis.

Leveraging the SEER Database to Understand Early Cancer Mortality at Scale: The Surveillance, Epidemiology, and End Results (SEER) database provides a unique opportunity to explore these questions with granular, population-based data spanning over two decades. With detailed information on tumor characteristics, demographics, staging, and initial treatments, SEER allows for in-depth analyses of patient trajectories from diagnosis onward, making it a powerful tool to study early cancer mortality patterns.

Study Objective: This study aims to investigate the determinants, interactions, and critical thresholds associated with early mortality in major cancer types such as Thoracic(Lung and Bronchus), Upper Digestive(Pancreas, Liver), Lower Digestive(Rectum, Ascending Colon, Sigmoid Colon) and breast cancer using SEER data from 2000 to 2022. Special emphasis will be placed on identifying high-risk subgroups and time-sensitive factors that could serve as targets for early intervention strategies.

Global importance: Although the data are drawn from the United States, the findings have broader relevance. Understanding the dynamics of short-term cancer survival can inform cancer control strategies in low- and middle-income countries (LMICs), where delays in diagnosis and treatment are often more pronounced. The results of this study may thus contribute to the development of global frameworks for timely and equitable cancer care.

LITERATURE REVIEW:

Short-Term Survival - A Blind Spot in Cancer Epidemiology: While 5-year survival is widely used as the gold standard to assess cancer outcomes, early mortality (typically defined as death within 3 to 12 months after diagnosis) remains underexplored. Yet, it captures a vulnerable window during which patients may die due to late diagnosis, aggressive disease, or treatment inaccessibility. Several studies, such as Exarchakou et al., 2018 and Elliss-Brookes et al., 2012, have shown that early deaths disproportionately affect socioeconomically disadvantaged groups, the elderly, and patients with advanced-stage cancers at diagnosis.

Determinants of Early Cancer Mortality: Early mortality is shaped by a complex interplay of tumor-related, patient-related, and system-related factors. Clinical determinants include cancer stage, grade, histology, and primary site. Patient-level determinants involve age, sex, race/ethnicity, comorbidities, and insurance status. System-level factors, such as diagnostic delays, access to specialized care, and initial treatment initiation, are increasingly recognized as pivotal in determining early outcomes (Neal et al., 2015; Lyratzopoulos et al., 2013).

Temporal Delays and Missed Opportunities: Diagnostic and therapeutic delays are critical contributors to poor short-term outcomes. The “patient interval” (symptom onset to presentation), “diagnostic interval” (first consultation to diagnosis), and “treatment interval” (diagnosis to treatment initiation) have each been associated with worse prognosis. While most evidence comes from high-income countries, similar patterns are observed or expected in LMICs, often with longer delays and lower system responsiveness (World Health Organization, 2017; Hanna et al., 2020).

Interactions and Risk Heterogeneity: Few studies have explored how the above determinants interact to influence risk dynamically. For example, the impact of age may vary depending on stage or cancer type; similarly, racial disparities may widen among patients with advanced disease. Understanding these interactions is crucial to identifying synergistic vulnerabilities and informing tailored interventions. Yet, large-scale studies using interaction-aware models (e.g., interaction terms in Cox models, machine learning approaches) remain scarce.

Gaps and Opportunities: Despite decades of research on cancer survival, thresholds that could signal high-risk situations (e.g., critical delays, age cutoffs, stage-specific risks) are rarely quantified. Identifying such thresholds can support triage, prioritization, and targeted intervention strategies. Studies using flexible modeling approaches—like splines, recursive partitioning, or threshold regression—are emerging, but remain limited. There is a clear need to complement existing work with data-driven identification of actionable time or risk cutoffs, particularly to inform policies in settings where resources are constrained.

SUMMARY: In sum, the literature points to a multidimensional yet fragmented understanding of early cancer mortality. While key determinants have been identified, their joint effects and critical thresholds remain poorly quantified. This study addresses that gap using a robust, population-based dataset and seeks to generate findings with global policy relevance.

SPECIFIC OBJECTIVES AND METHODOLOGY:

Objective 2: To identify demographic and clinical determinants of short-term survival.

Hypothesis (H2): Older age, advanced stage at diagnosis, and certain cancer types are significantly associated with lower short-term survival.

Methodology:

Variables: - Outcome: Binary survival status at 1, 3, and 6 months. - Predictors: Age, sex, race/ethnicity, stage, grade, cancer site, year of diagnosis. - Statistical model: * Multivariable logistic regression for each time point (1-, 3-, and 6-month STS). Adjusted Odds Ratios (aORs) with 95% confidence intervals. Model validation: Hosmer–Lemeshow test, AUC, and calibration plots.

Objective 3: To explore interactions between predictors of early mortality.

Hypothesis (H3): Certain risk factors (e.g., advanced age and late stage) may have synergistic effects on early mortality.

Methodology: - Inclusion of interaction terms (e.g., age × stage, stage × cancer type) in the logistic regression models. - Likelihood ratio tests to compare models with and without interaction terms. - Marginal effect plots or interaction plots for interpretation.

Objective 4: To identify clinical thresholds or cutoffs associated with abrupt changes in short-term survival probability.

Hypothesis (H4): Non-linear thresholds (e.g., age > 75 or stage IV) exist where survival probability drops significantly.

Methodology: - Use of Classification and Regression Trees (CART) to detect data-driven cutoffs. - Complementary approach with segmented (piecewise) regression to identify inflection points. - Bootstrap validation to ensure robustness of thresholds.

Objective 5: To compare cancer-specific vs. non-cancer mortality during the first 6 months post-diagnosis.

Hypothesis (H5): A substantial proportion of early deaths are not due to cancer itself, especially among older or comorbid patients.

Methodology: - Classification of deaths as cancer-specific vs. non-cancer using SEER cause of death variable. - Competing risks analysis using the Fine and Gray model: - Subdistribution hazard ratios (SHRs) for cancer-specific mortality. - Stratification by cancer type and stage. - Cumulative incidence curves to illustrate competing risks over time.

Model Diagnostics

For all statistical models, appropriate diagnostics will be conducted to ensure validity and robustness.

Model Diagnostics Details
Logistic Regression (Objectives 2 and 3) Discrimination ROC Curve, AUC (>0.7 acceptable, >0.8 good)
Calibration Calibration plot (predicted vs observed), Hosmer–Lemeshow test (non-significant p-value = good fit)
Multicollinearity Variance Inflation Factor (VIF > 5 or > 10 indicates multicollinearity)
Influential Observations Cook’s Distance, Leverage, DFBETAs
Linearity of the Logit Box-Tidwell test or residual plots
Interaction Terms Interpretation Plot marginal effects
Model Comparison Likelihood Ratio Test (LRT) for models with vs without interaction terms
Classification and Regression Trees (CART) (Objective 4) Overfitting Check Cross-validation (e.g., 10-fold CV), prune the tree
Complexity Parameter Selection Based on minimizing cross-validated error (cp)
Stability Check Bootstrap sampling
Competing Risks Model (Fine and Gray) (Objective 5) Proportionality Assumption Test time-varying effects (interaction with time), Schoenfeld residuals (adapted)
Model Fit Compare observed vs predicted cumulative incidence curves

ANALYSIS AND RESULTS:

Data cleaning and pre-processing:

In order to ensure data quality, we performed several cleaning steps. - We first renamed the variables for better clarity.
- We filtered the dataset to retain only selected major cancer types and regrouped them by creating a new variable that classified anatomical subsites into broader cancer categories.
- The survival_months variable was converted to numeric, and observations with missing survival times were removed.
- We also created three new binary indicators (death_within_1m, death_within_3m, and death_within_6m) to capture cancer-specific death within 1, 3, and 6 months after diagnosis.
- Finally, the dataset was arranged by cancer type and year of diagnosis to improve readability and facilitate subsequent analyses.

Click here to access the data cleaning script.

cleaned dataset description:

The dataset includes cancer patients diagnosed over multiple years, with detailed demographic, clinical, and treatment information. Key variables cover age at diagnosis, sex, race, marital status, cancer site and group, tumor characteristics (size, stage, grade), treatments received (surgery, radiation, chemotherapy), and survival outcomes. Early mortality within 1, 3, and 6 months is also captured, along with cause-specific death indicators. Observations are organized chronologically by year of diagnosis, cancer site, and age group.