Chest X-ray AI Model Predictions and Fairness Analysis

Processed data from 2026 accompanies a study on AI performance disparities in chest X-ray analysis. It contains per-image pneumonia prediction scores from five DenseNet-121 models evaluated on the RSNA Pneumonia Detection Challenge dataset, along with view type metadata and analysis outputs. The dataset supports research into how technical acquisition parameters, such as radiograph view type, influence model fairness compared to demographic factors.

Use Cases

• Analyzing AI performance disparities based on technical parameters like AP/PA view type.
• Validating fairness metrics, such as Cohen's d effect sizes, across different model architectures.
• Conducting intersectional analysis of model bias using pre-computed ANOVA decomposition outputs.
• Benchmarking pneumonia detection models against a standardized external validation framework.

Strengths

• Includes predictions from five distinct DenseNet-121 models, enabling comparative analysis.
• Contains formal statistical analysis outputs like ANOVA decomposition and Cohen's d effect sizes.
• Dataset is cross-platform and associated with a peer-reviewed PLOS Digital Health manuscript.

Limitations

• Raw DICOM images are not included; users must source them separately from RSNA or NIH.
• The exact number of rows is unclear; one source states 'n=26,684' while another lists 'Rows: None'.
• Column names and full data schema are not explicitly provided in the available metadata.

Provenance

Source

Processed from the RSNA Pneumonia Detection Challenge dataset.

Collection Method

Generated as analysis outputs for a specific research study.

Freshness

Last updated on 2026-03-18.