Developing a clinical utility framework to evaluate prediction models in radiogenomics.

Area of Science:

• Radiogenomics
• Medical Imaging
• Genetics
• Biostatistics

Background:

• Radiogenomics, integrating imaging and genetic data, is an emerging field for disease prediction.
• Established methodologies for evaluating radiogenomics prediction models are lacking.
• Breast cancer risk assessment can benefit from advanced predictive modeling.

Purpose of the Study:

• To develop and assess a clinical decision framework for evaluating radiogenomics prediction models in breast cancer.
• To compare the predictive performance of models using Gail risk factors, single nucleotide polymorphisms (SNPs), and BI-RADS mammographic features.
• To determine the utility of integrating genetic and imaging data for enhanced breast cancer risk estimation.

Main Methods:

• A retrospective case-control study collected Gail model risk factors, SNPs, and BI-RADS lexicon features.
• Three logistic regression models were constructed: Gail, Gail+SNP, and Gail+SNP+BI-RADS.
• Receiver Operating Characteristic (ROC) curves, utility analysis, Maximum Expected Utility (MEU), and McNemar's test were employed for model evaluation.

Main Results:

• Integrating SNPs and BI-RADS features improved the Area Under the ROC Curve (AUC) and MEU compared to the baseline Gail model.
• SNPs increased model sensitivity (0.276 vs. 0.147) but decreased specificity (0.855 vs. 0.912).
• Adding mammographic features further enhanced sensitivity (0.457) and specificity (0.872), with both SNPs and mammographic features showing significant impact (p < 0.001).

Conclusions:

• The developed decision framework, incorporating utility analysis and McNemar's test, offers a novel approach for evaluating radiogenomics prediction models.
• Genetic variants (SNPs) and mammographic features significantly enhance breast cancer risk prediction accuracy.
• This framework facilitates the robust assessment of radiogenomics techniques for clinical decision-making.