Evaluating Sociodemographic Biases in Artificial Intelligence-Based Glioblastoma Response Assessment Algorithms

Area of Science:

• Artificial Intelligence in Medical Imaging
• Neuro-oncology
• Radiomics and Quantitative Imaging

Background:

• Artificial intelligence (AI) models in medical imaging may exhibit biases, but the underlying causes are not fully understood.
• This study investigates sociodemographic biases in AI-based glioblastoma MRI segmentation.
• Four nnUNet models were trained on datasets varying in size and demographic composition.

Purpose of the Study:

• To evaluate potential sociodemographic biases in AI glioblastoma MRI segmentation models.
• To assess the impact of training dataset size and demographic composition on AI model performance.
• To identify factors influencing bias in AI-driven tumor segmentation.

Main Methods:

• Four AI models (FeTS2, BraTS 2024, small homogenous, small heterogenous) were evaluated.
• An independent dataset of 480 patients from a single academic center was used for bias assessment.
• Automated segmentations (FLAIR, enhancing tumor) were scored using Dice scores; beta regression analyzed sociodemographic influences.

Main Results:

• The model trained on a homogenous dataset (White, non-Hispanic males) showed the lowest Dice scores and significant age/smoking status biases.
• The BraTS 2024 model achieved the highest Dice scores (0.996 FLAIR, 0.999 Enhancement) with minimal bias.
• Demographic heterogeneity in training data correlated with reduced bias, irrespective of dataset size.

Conclusions:

• Sociodemographic bias in glioblastoma MRI segmentation AI is generally low.
• AI models trained on smaller, homogenous datasets exhibit greater bias.
• Increased demographic heterogeneity in training data, even without larger datasets, reduces AI bias, as exemplified by the BraTS model's superior performance.