Rethinking fairness in AI to improve current practice in oncology.

Area of Science:

• Oncology
• Artificial Intelligence
• Health Equity

Background:

• Fairness in artificial intelligence (AI) is crucial, especially in oncology.
• Existing fairness metrics are often flawed and do not account for real-world inequities.
• Patient diversity and structural biases significantly impact cancer care outcomes.

Purpose of the Study:

• To highlight the limitations of current fairness metrics in AI for oncology.
• To emphasize the need for revised fairness frameworks that address biases in data and predictions.
• To advocate for improved equity in cancer care through better AI evaluation.

Main Methods:

• Analysis of biases in ground truth labels, AI predictions, and demographic attributes.
• Critique of standard fairness metrics in the context of oncology.
• Conceptual framework development for equitable AI in cancer care.

Main Results:

• Current fairness metrics in AI for oncology are inadequate.
• Biases are present in datasets, predictions, and demographic data, distorting fairness assessments.
• A significant gap exists between current AI fairness evaluations and true equity in cancer care.

Conclusions:

• Rethinking AI fairness frameworks is essential for oncology.
• Addressing biases in data and metrics is critical for achieving health equity.
• New approaches are needed to ensure AI promotes equitable cancer care outcomes.