Model Comparison for Breast Cancer Prognosis Based on Clinical Data

Area of Science:

• Oncology
• Biostatistics
• Machine Learning in Medicine

Background:

• Accurate breast cancer prognosis is crucial for effective treatment planning.
• Numerous predictive models exist, but their comparative performance for breast cancer prognosis requires further investigation.

Purpose of the Study:

• To compare the performance of eight diverse predictive models for breast cancer prognosis.
• To identify key clinical predictors influencing breast cancer prognosis across different timeframes.

Main Methods:

• Utilized a dataset of 1,981 breast cancer patients, retaining 11 key clinical predictors.
• Evaluated eight models: Generalized Linear Model (GLM), GLM-Net, Partial Least Square (PLS), Support Vector Machines (SVM), Random Forests (RF), Neural Networks, k-Nearest Neighbors (k-NN), and Boosted Trees.
• Employed paired t-tests on data resampling to compare model performance using Area Under the ROC curve (AU-ROC).

Main Results:

• Random Forests, Boosted Trees, PLS, and GLM-Net demonstrated slightly superior overall performance compared to other models.
• Identified key predictors of breast cancer prognosis: number of positive lymph nodes, tumor size, cancer grade, and estrogen receptor status.
• Analyzed the differential short-term and long-term impact of clinical variables on prognosis.
• Combination of chemotherapy and radiotherapy showed the most significant impact on breast cancer prognosis among treatment plans.

Conclusions:

• While several models show promising results, the performance differences are marginal, suggesting robustness in breast cancer prediction.
• Clinical factors like lymph node status, tumor size, grade, and hormone receptor status are critical for accurate prognosis.
• Treatment modalities, particularly combined chemo/radiotherapy, play a substantial role in influencing long-term breast cancer outcomes.