Aasa Feragen
Measuring the reliability of trustworthy healthcare AI for mental health risk prediction
Artificial intelligence (AI) is increasingly used as a tool both for screening, diagnosis and resource allocation in healthcare. Trustworthy AI solutions like algorithmic fairness and explainable AI are often used to improve the safety and reliability of such AI solutions.
However, in cases where the ground truth diagnoses used to train and test AI have demographically dependent biases, explainability and algorithmic fairness can fail to provide the expected protection. This can happen if specific demographic groups are systematically under- or overdiagnosed.
This project takes risk prediction for major depressive disorder as a use case and aims both to clarify and showcase the potential failure mechanisms - and to provide technical tools to avoid them.
