Minimum sample size calculation for radiomics-based binary outcome prediction models: Theoretical framework and practical example

Abstract

Background and purpose

Determining the appropriate sample size for developing robust radiomics-based binary outcome prediction models and identifying the maximum number of predictors safely allowable within a fixed dataset size remain critical yet challenging tasks. This study aims to propose and demonstrate a structured method for addressing these issues, enhancing methodological rigor and practicality in radiomics research.

Materials and methods

We introduce a comprehensive sample size calculation framework for binary outcome prediction models in radiomic studies. The proposed approach integrates three key criteria: (1) maintaining a global shrinkage factor (S) ≥ 0.9 to control model overfitting, (2) ensuring a minimal absolute difference between apparent and adjusted performance metrics, and (3) precisely estimating the overall outcome risk. Additionally, we develop an accessible online calculation tool enabling researchers to efficiently determine either the minimum sample size or the maximum number of predictors permissible, based on clearly defined statistical parameters.

Results

The presented method systematically addresses model overfitting by integrating a global shrinkage factor into the calculation, providing robust estimates compared with traditional heuristic approaches (“rules of thumb”). Practical examples demonstrate that this structured method effectively balances predictive accuracy and generalizability, while the online tool provides researchers with a user-friendly platform to perform the necessary calculations.

Conclusion

Clear justification of sample size decisions is essential for developing reliable predictive models in radiomics research. By adopting a structured and rigorous calculation method, researchers can effectively minimize overfitting, ensure accurate risk estimation, and substantially enhance the reliability and validity of their predictive models.