Machine Learning and Feature Selection Applied to SEER Data to Reliably Assess Thyroid Cancer Prognosis

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Historical clinical datasets offer valuable insights for improving patient care and physician decision-making.
• Effective utilization of existing medical records can enhance predictive capabilities in oncology.

Purpose of the Study:

• To develop and validate machine learning models for predicting thyroid cancer patient prognosis.
• To assess the accuracy of feature selection algorithms and supervised neural networks in analyzing clinical data for survival prediction.

Main Methods:

• Analysis of a SEER database containing clinical features from thyroid cancer patients.
• Application of feature selection algorithms: Fisher's discriminant ratio, Kruskal-Wallis' analysis, and Relief-F.
• Optimization of supervised neural networks, specifically multilayer perceptrons, for patient survival prediction.

Main Results:

• Achieved 94.5% accuracy in distinguishing between thyroid cancer patients with different survival outcomes (over 10 years vs. less than 5 years).
• Demonstrated the reliability of using unspecialized medical recordings for predictive modeling.
• Successfully identified key clinical variables for prognosis prediction.

Conclusions:

• Machine learning, combined with feature selection, provides a powerful tool for predicting thyroid cancer patient prognosis.
• This strategy can reliably transform existing medical data into actionable predictive insights for optimized treatment decisions.
• The developed machine learning approach holds potential for application in other diseases and clinical trial design.