Background

In cardiology, the use of artificial intelligence (AI) has been proposed to diagnose cardiovascular disease (CVD) from cardiac magnetic resonance imaging data [1], or other data sources such as electrocardiograms [2] and clinical variables. Some work has sought to produce “multimodal” AI models that can interpret and analyse multiple such data sources [3]. Multimodal AI models have great potential as they can exploit complementary information from different sources in the same way as clinicians do when making decisions.

Recently, there has been increasing concern about the potential of AI models to exhibit “biased” behaviour. For example, AI models trained using mostly data from a single demographic group (such as White males) may perform well for that group but less well on other groups, i.e. the model can be biased or unfair [4]. In medicine, unfair AI models have the potential to impact patient outcomes disproportionately, maintaining or even exacerbating healthcare disparities. 

Novelty & Importance:

So far, work on fairness in AI in medicine has focused on single modality AI models. Multimodal AI models have great potential across a range of applications but they are also probably as susceptible to bias as single modality AI models. It is important to investigate this possibility as AI models start to be translated into the clinic and have impact on patients in the real world.

Aim:

This project aims to investigate the hypothesis that multimodal AI models for CVD characterisation can exhibit biased behaviour and if appropriate develop solutions to address this bias.

References:

[1] Bernard, et al. """"Deep Learning Techniques for Automatic MRI Cardiac Multi-structures Segmentation and Diagnosis: Is the Problem Solved?"""" IEEE Transactions on Medical Imaging, 2018 (https://doi.org/10.1109/TMI.2018.2837502)

[2] Attia et al. “Application of Artificial Intelligence to the Electrocardiogram,” European Heart Journal, 2021 (https://doi.org/10.1093/eurheartj/ehab649)

[3] Acosta et al. “Multimodal Biomedical AI,” Nature Medicine, 2022 (https://doi.org/10.1038/s41591-022-01981-2)

[4] Puyol-Antón et al. “Fairness in Cardiac Magnetic Resonance Imaging: Assessing Sex and Racial Bias in Deep Learning-Based Segmentation,"""" Frontiers in Cardiovascular Medicine, 2022 (https://doi.org/10.3389/fcvm.2022.859310)

"""