Benchmarking prognosis methods for survivability - A case study for patients with contingent primary cancers

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Growing incidence of second primary cancers necessitates robust survival prediction methods.
• Current prognosis models for multiple primary cancers lack comprehensive benchmarking.

Purpose of the Study:

• To evaluate the five-year survival prognosis performance of six machine learning models.
• To identify optimal models for predicting survivability in patients with multiple primary cancers.

Main Methods:

• Six machine learning approaches were assessed: artificial neural network, decision tree (DT), logistic regression, support vector machine, naïve Bayes (NB), and Bayesian network (BN).
• Synthetic minority over-sampling technique (SMOTE) was employed to address data imbalance.
• A Taiwanese nationwide cancer patient database (7,845 subjects) was utilized, focusing on ten primary and secondary cancer types.

Main Results:

• Models incorporating SMOTE demonstrated significant improvements in sensitivity and specificity.
• Naïve Bayes (NB) exhibited the highest accuracy and specificity.
• Bayesian Network (BN) achieved the highest sensitivity, while NB, BN, and DT excelled in computational time and knowledge representation.

Conclusions:

• Appropriate machine learning models can enhance the precision of survival predictions for patients with multiple primary cancers.
• Accurate prognosis models can significantly support clinical decision-making and treatment recommendations.