Predicting liver cancers using skewed epidemiological data.

Area of Science:

• Oncology
• Data Science
• Machine Learning

Background:

• Liver cancer poses a significant global health risk.
• Early identification of high-risk individuals is crucial for effective intervention.
• Machine learning offers a cost-effective approach for cancer prediction.

Purpose of the Study:

• To address the challenge of class imbalance in machine learning models for liver cancer prediction.
• To introduce and evaluate novel undersampling techniques for skewed medical datasets.
• To improve the reliability and accuracy of liver cancer forecasting models.

Main Methods:

• Systematic evaluation of existing class-imbalance solutions.
• Introduction of two novel undersampling methods: K-means++ based and Learning Vector Quantization (LVQ) based.
• Application of the developed algorithm to a five-year liver cancer prediction dataset from China.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.76 using only epidemiological data.
• Demonstrated superior performance compared to oversampling, existing undersampling, ensemble methods, and outlier detection algorithms.
• Successfully rebalanced positive and negative samples, mitigating model bias.

Conclusions:

• The proposed undersampling methods provide a feasible and practical solution for handling skewed medical data in cancer prediction.
• This approach enhances the potential for individualized interventions and improved patient outcomes.
• The study offers a roadmap for developing more accurate and reliable predictive models in oncology.