Supervised discretization can discover risk groups in cancer survival analysis

Area of Science:

• Oncology
• Medical Informatics
• Biostatistics

Background:

• Discretization of continuous variables is crucial for identifying patient risk groups in medical research.
• Traditional categorization methods, such as the TNM+A protocol, are standard but may have limitations in precision.

Purpose of the Study:

• To compare the performance of supervised machine learning discretization methods (CAIM, ChiM, DTree) against the gold-standard TNM+A protocol for breast cancer patient stratification.
• To evaluate the clinical utility of alternative discretization algorithms in cancer patient data analysis.

Main Methods:

• Employed three supervised machine learning discretization algorithms: CAIM, ChiM, and DTree.
• Utilized standard survival analysis techniques, including Kaplan-Meier curves, Cox regression, and predictive modeling, to assess discretization performance.
• Analyzed patient data from the Medical Oncology Service of Hospital Clínico Universitario (Málaga, Spain), with a follow-up period from 1982 to 2008.

Main Results:

• Supervised discretization algorithms demonstrated comparable or superior performance to the TNM+A protocol in patient stratification.
• The application of CAIM, ChiM, and DTree provided valuable insights into breast cancer diagnosis and patient outcomes.
• Survival analysis revealed significant differences in risk stratification based on the applied discretization methods.

Conclusions:

• Alternative discretization algorithms from machine learning offer a powerful approach for stratifying breast cancer patients.
• These methods can provide clinicians with enhanced information for improved diagnosis and prognosis.
• The study highlights the potential of machine learning in refining patient stratification for better cancer care.