Exploring Artificial Intelligence Biases in Predictive Models for Cancer Diagnosis.

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence

Background:

• The American Society of Clinical Oncology (ASCO) established principles for responsible artificial intelligence (AI) use in oncology.
• The adherence to these principles in published research remains largely unassessed.

Purpose of the Study:

• To evaluate the presence of biases and the quality of studies on AI models for cancer diagnosis.
• To assess compliance with ASCO's responsible AI principles.
• To examine the impact of these factors on subsequent research applications.

Main Methods:

• A systematic review of AI predictive models for cancer diagnosis published in an ASCO informatics journal.
• Evaluation using 17 bias criteria aligned with ASCO principles and the CREML checklist for study quality.
• Analysis of performance metrics and citation counts.

Main Results:

• Nine studies were included, revealing common biases: environmental, life-course, contextual, provider expertise, and implicit bias.
• Transparency, oversight, privacy, and human-centered AI application were the least adhered-to ASCO principles.
• Only 22% of studies provided data access, and CREML checklist indicated methodological and reporting deficiencies.

Conclusions:

• Most AI oncology studies exhibit biases and reporting flaws, limiting their applicability and reproducibility.
• There is a significant gap in adherence to ASCO's responsible AI principles.
• Recommendations include enhancing transparency, data accessibility, and adherence to international guidelines for reliable AI research in oncology.