3615 - Development and Validation of a Machine Learning-Based Size-Only Staging System for Improved Prognostic Stratification in Ten Common Cancers

Purpose/Objective(s):

The study addresses limitations in the American Joint Committee on Cancer (AJCC) tumor (T) staging system, particularly the lack of significant differences in overall survival (OS) hazard ratios (HRs) and overlapping survival curves across certain stages. A novel size-only staging (SOS) system, utilizing machine learning-based clustering (MLC), was developed to improve T staging accuracy.

Materials/Methods:

The SOS system was trained across 961,600 patients with ten common cancers using data from the Surveillance, Epidemiology, and End Results (SEER) database. The study employed hierarchical clustering, a machine learning technique, to categorize tumors into distinct stages based on tumor size and patient survival data. Hierarchical clustering was chosen for its ability to build a hierarchy of clusters by successively merging or splitting data points, using Euclidean distance as the similarity measure and Ward's method for linkage to minimize variance within clusters. The clustering process was applied to tumor size data, and the resulting clusters were validated using Cox proportional hazards modeling, log-rank tests, and Kaplan-Meier survival analyses. The three-class SOS (SOS-3) and four-class SOS (SOS-4) systems were evaluated for their ability to resolve issues of insignificant HR differences and overlapping survival curves. Data preprocessing included outlier detection using the interquartile range (IQR) method to ensure robustness.

Results:

The SOS-3 system resolved issues of insignificant HR differences between certain stages, such as T1 and T4 in prostate cancer (HR for T1 vs. T4: 0.93, p=0.52) and T1 and T2 in esophageal cancer (HR for T1 vs. T2: 1.13, p=0.07). The SOS-4 system further improved OS distinctions, minimizing crossover survival curve issues observed in AJCC staging, particularly in prostate, bladder, and stomach cancers. For example, in prostate cancer, the SOS-4 system introduced a new category for tumors sized 1-7 mm, with a 10-year survival rate of 79.6% (95% CI: 78.2-80.9), compared to 86.1% (95% CI: 85.1-87.1) for tumors sized 15-22 mm. Significant differences in HRs were observed across most stages in both SOS-3 and SOS-4 systems, with p-values <0.005 for the majority of comparisons.

Conclusion:

The SOS system offers a refined approach to T staging, providing more accurate prognostic distinctions across multiple cancer types. This innovative system has the potential to enhance imaging protocols, prognostic factors, and clinical trial designs, paving the way for more personalized and effective cancer treatment strategies.