Measuring the reliability of trustworthy healthcare AI for mental health risk prediction

Trustworthy artificial intelligence (AI) methods aim to increase the fairness and transparency of AI algorithms. While this is a noble goal, current approaches are not well understood and actually risk making matters worse: Enforcing algorithmic fairness in diagnostic AI where some groups are systematically under-diagnosed can lead to increased bias in algorithms. Unfortunately, mental health is notorious in the sense that correct diagnosis depends on individuals' willingness to seek care and openly describe their symptoms -- qualities that have a strong demographic and cultural dependency. Additionally, naive interpretation of explainable AI (XAI) feedback can lead to incorrect conclusions on which features are important. This is particularly problematic in applications where there is a demographic bias, and the algorithm might be using demographic information to recalibrate the algorithm rather than to actually discover disease. Such incorrect feedback risks both leading to incorrect scientific discovery about diseases, and decreasing the users' trust in the algorithm. This project aims to bring these problems out in the open: By using major depressive disorder (MDD) risk scoring algorithms as a use case, we will illustrate how different levels of label bias affect trustworthy AI algorithms; we will estimate the magnitude of the bias and illustrate how it can be accounted for.

This project is anchored in the group of Professor Aasa Feragen at DTU Compute, Denmark’s Technical University and carried out in collaboration with Melanie Ganz from the Neurobiological Research Unit at Rigshospitalet, Denmark.

Project participants:

• Professor Aasa Feragen (PI), DTU Compute, Denmark's Technical University (DTU)
• Associate Professor Melanie Ganz-Benjaminsen, Neurobiology Research Unit, Copenhagen University Hospital/Rigshospitalet
• Postdoc Olalekan Joseph Akintande, DTU Compute, Denmark's Technical University (DTU)
• PostDoc Helen Coupland, Neurobiology Research Unit, Copenhagen University Hospital/Rigshospitalet