An approach for cancer outcomes modelling using a comprehensive synthetic dataset

Area of Science:

• Medical informatics
• Machine learning in oncology
• Radiomics and clinical data integration

Background:

• Limited patient data hinders machine learning (ML) model development for cancer outcome prediction.
• Existing synthetic image generation methods often lack crucial prognostic information.

Purpose of the Study:

• To develop a cancer outcome modeling approach using a comprehensive synthetic dataset that accurately mimics real patient data.
• To evaluate the performance of ML models trained on synthetic data for predicting patient survival.

Main Methods:

• A real dataset of 132 non-small cell lung cancer patients (CT-based radiomic and clinical features) was utilized.
• A synthetic dataset was generated using a conditional tabular generative adversarial network.
• Models predicting two-year overall survival were trained using various feature selection methods and ML algorithms, then tested on real data.

Main Results:

• Real and synthetic datasets showed high similarity (average 1-minus-KS statistic of 0.871 for continuous features; p<0.001 for discrete features).
• XGBoost with random forest importance features demonstrated consistent performance across both datasets (<1.3% difference in balanced accuracy and AUC-PR).

Conclusions:

• Synthetic radiomic and clinical data augmentation shows potential for cancer outcome modeling.
• Further validation with larger, diverse datasets is essential for broader application beyond lung cancer.