Representation and Bias in Artificial Intelligence Models for Thyroid Cancer: A Systematic Review

Area of Science:

• Medical Informatics
• Oncology
• Artificial Intelligence

Background:

• Growing interest in artificial intelligence (AI) for thyroid cancer care, aiming to improve diagnostics, predict outcomes, and personalize treatment.
• Potential for AI bias in thyroid cancer algorithms, leading to disparities in care for underrepresented populations.
• Need for systematic evaluation of demographic representation and bias in AI models for thyroid cancer.

Purpose of the Study:

• To systematically review AI models used in thyroid cancer care.
• To assess demographic representation and identify potential biases within these AI models.
• To evaluate the alignment of AI model data with global thyroid cancer epidemiology.

Main Methods:

• Systematic literature search on EMBASE, PubMed, and Google Scholar up to January 2024.
• Inclusion of studies involving AI for thyroid cancer management with demographic data.
• Data extraction and risk-of-bias assessment by two independent reviewers, registered on PROSPERO.

Main Results:

• 197 studies reviewed, primarily focusing on diagnosis (133) and prediction/prognosis (47).
• Predominant participant origins from China (124) and the US (26); more females (12,410) than males (4222).
• Significant underrepresentation of East Asians and overrepresentation of White and Black participants compared to global prevalence; socioeconomic factors often omitted.

Conclusions:

• Current AI models for thyroid cancer exhibit significant gaps in data diversity and representativeness.
• AI models align with general epidemiological trends but lack comprehensive demographic inclusion.
• Development of more equitable AI models, considering diverse demographics and sociocultural backgrounds, is crucial for improving global thyroid cancer care.