De-biased and interpretable foundation models to reconstruct health history from fundus images in The Gambia

• Applicant: Professor Dr. Philipp Berens
• Subject Area: Medical Informatics and Medical Bioinformatics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 571331899

Despite substantial progress, sub-Saharan Africa is still lagging behind other world regions in major health outcomes and faces both persisting health challenges through infectious diseases as well as increasingly through non-communicable diseases such as cardiovascular conditions and diabetes. Artificial intelligence (AI) has been claimed to hold great promise for improving health outcomes in low-and-middle income countries and for alleviating health inequalities. For example, low-cost devices powered by deep learning algorithms have been proposed to mitigate the issue that an insufficient number of trained specialists are available e.g. to perform effective screening for common diseases such as diabetic retinopathy. However, most AI systems show potential for bias, as they are typically trained on data from the Global North, since only few well-annotated high-quality datasets are available from sub-Saharan Africa. Therefore, biases introduced during training may propagate through the entire life cycle of AI systems and they may perform less well than believed. This problem is made even more pressing by a new class of models called foundation models, which have been trained in the Global North on millions of images using vast resources. Here, we aim to evaluate algorithmic strategies to de-bias existing foundation models in ophthalmology and develop techniques to make such models interpretable. To this end, we will collect a large database of fundus images from participants of an existing health cohort in The Gambia using mobile fundus cameras. This health cohort will give us a unique chance to link the fundus images to the health history of the participants. At the same time, we will develop supervised and unsupervised methods to assess and remove the bias in the representations provided by existing foundation models and develop algorithmic strategies to make foundation models inherently interpretable. Finally, we will use the collected data, the fine-tuned, debiased foundation model and the information available in the health cohort to implement an oculomics approach for reconstructing health history from the fundus images. Here, our hypothesis is that it is possible to infer information about prior severe episodes of infectious diseases such as malaria or tuberculosis from fundus images in addition to health status related information such as hypertension, which has been established to work well using modern deep learning. Together, this will provide us with a unique African fundus image dataset to study the potential of a cheap and non-invasive technique to obtain health information about individuals as well as validated machine learning algorithms to study and overcome potential biases in medical foundation models.

• DFG Programme: Research Grants
• International Connection: Gambia
• International Co-Applicant: Professor Bubacarr Bah, Ph.D.