The potential of GPT-4 advanced data analysis for radiomics-based machine learning models

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Oncology
• Radiomics and Computational Pathology

Background:

• Glioma molecular subtyping is crucial for treatment decisions.
• Radiomics analysis of MRI data offers potential for non-invasive glioma classification.
• Developing automated machine learning models (MLMs) can streamline this process.

Purpose of the Study:

• To evaluate GPT-4's Advanced Data Analytics (ADA) package for autonomous MLM development.
• To predict glioma molecular types using radiomics features from MRI.
• To benchmark ADA's performance against established handcrafted MLMs.

Main Methods:

• Radiomic features extracted from 615 preoperative MRIs.
• Multiclass ML approach using ADA to build an ML pipeline.
• Comparison with a handcrafted model using N4, Zscore, and WhiteStripe normalization.
• External validation on two independent glioma datasets (D2, D3).

Main Results:

• GPT-4 achieved 0.820 accuracy on D3, significantly outperforming the benchmark (0.678).
• GPT-4 showed high recall for IDH-wildtype glioma (0.997) but lower recall for IDH-mutant types.
• Performance on the D2 dataset was lower than the benchmark, with similar class-wise variations.

Conclusions:

• GPT-4's ADA can autonomously develop competitive radiomics-based MLMs.
• Class-wise performance disparities indicate limitations in handling imbalanced datasets.
• Further refinement is needed for robust end-to-end ML pipeline automation in glioma subtyping.