Demographic distribution matching between real-world and virtual phantom population.

Area of Science:

• Medical Imaging
• Biostatistics
• Clinical Trials

Background:

• Virtual imaging trials (VITs) offer cost-effective alternatives to traditional clinical trials.
• Demographic disparities between virtual and real cohorts can bias imaging performance assessments.
• Unaddressed biases limit the translational relevance of VIT findings for diverse patient groups.

Purpose of the Study:

• Introduce DISTINCT (Distributional Subsampling for Covariate-Targeted Alignment), a statistical framework for aligning demographic subsamples from clinical datasets with virtual cohorts.
• Enable robust comparisons between virtual imaging trial populations and real-world clinical data.

Main Methods:

• Applied DISTINCT to the National Lung Screening Trial (NLST) and a virtual dataset (VLST).
• Jointly aligned continuous (age, BMI) and categorical (sex, race, ethnicity) variables using multidimensional bins.
• Evaluated demographic similarity using Wasserstein and Kolmogorov-Smirnov distances.
• Assessed lung cancer risk prediction performance and stability using ROC analysis on aligned subsamples.

Main Results:

• DISTINCT identified a maximal demographically aligned NLST subsample of 9974 participants matching the VLST population.
• Area under the curve (AUC) estimates for lung cancer risk scores stabilized around 6000 participants.
• Stratified analyses highlighted demographic-specific variations in AUC, emphasizing the need for covariate alignment.

Conclusions:

• DISTINCT offers a statistically rigorous and scalable method for covariate alignment between real and virtual imaging cohorts.
• The framework is applicable to various imaging modalities, diseases, and factors of variability.
• DISTINCT facilitates fair and representative performance assessments, advancing VIT integration in research and protocol optimization.