Ensuring Fairness in Detecting Mild Cognitive Impairment with MRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Machine learning (ML) is vital for Alzheimer's Disease (AD) diagnosis, but current methods struggle with Mild Cognitive Impairment (MCI) detection.
• Challenges in MCI classification include label imbalance and bias from sensitive attributes in neuroimaging data.

Purpose of the Study:

• To develop an end-to-end, fairness-aware ML approach for label-imbalanced MCI classification using neuroimaging data.
• To enhance the accuracy and equity of MCI detection, a critical stage preceding AD.

Main Methods:

• An end-to-end fairness-aware classification approach was designed, integrating the FACIMS framework into the STREAMLINE automated ML environment.
• The method specifically addresses label imbalance and bias in neuroimaging datasets for MCI classification.
• Performance was evaluated against nine other ML algorithms.

Main Results:

• The proposed fairness-aware approach achieved balanced accuracy comparable to existing methods.
• The method demonstrated a prioritization of fairness across five different sensitive attributes in MCI classification.
• This represents a significant step towards equitable ML diagnostics for MCI.

Conclusions:

• The developed fairness-aware ML approach effectively addresses label imbalance and bias in MCI detection from neuroimaging data.
• This method offers a more equitable and reliable tool for early identification of MCI, crucial for AD progression management.
• The integration into STREAMLINE facilitates automated and fair ML diagnostics in clinical settings.