A downsampling strategy to assess the predictive value of radiomic features

Area of Science:

• Radiomics
• Medical Imaging Analysis
• Machine Learning in Healthcare

Background:

• Radiomic models are crucial for prediction tasks, but model failure offers little insight into data relevance.
• Distinguishing between insufficient data and irrelevant features is challenging when radiomic models underperform.

Purpose of the Study:

• To propose and validate a downsampling method for assessing radiomic model performance in two-group classification.
• To differentiate between data scarcity and feature irrelevance as causes for poor model performance.

Main Methods:

• Utilized two large patient cohorts to create experimental configurations with varying patient numbers.
• Developed univariate and multivariate radiomic models from each configuration.
• Compared model performance (Youden Index and Area Under the Curve) against stable performance from maximum patient data.

Main Results:

• Multivariate, machine learning-based radiomic models showed improved performance with increased patient numbers.
• Univariate models demonstrated decreased performance as patient numbers varied.
• The downsampling method accurately predicted achievable Youden Index and Area Under the Curve with larger datasets.

Conclusions:

• The downsampling method effectively estimates the potential performance of radiomic models with more data.
• This approach can identify when radiomic features lack relevant information for a specific classification task.