Narges Razavian, an associate professor at the New York University Langone Health, provided a keynote for the NLDL conference 2024 (11 Jan 2024).
Title: Fair Self supervised Learning in multiple modalities (Imaging, EHR, and in combination) with Applications to Medicine
Abstract: Recent progress in self-supervised learning (SSL), and availability of large clinical datasets that include millions of records of medical imaging and electronic health records (EHRs) provide an untapped opportunity to improve representation learning that is crucial to almost all medical predictive models. Learning fair and strong representations via SSL is still an under-explored and challenging problem, and this issue is specially important in medicine, where there is an imbalance in amount of data available for each sub-populations. Additionally, working with diverse modalities in medicine (EHRs, 3D imaging, Large Histopathology images) impose several unsolved challenges. In this talk, I will share a collection of recent work from my team, addressing many of the methodological shortcomings of state of the art methods. Specifically, I will present novel methods to improve self supervised learning in medical imaging and EHR settings; Novel methods to improve fairness of the SSL trained method; and we will end by discussing future directions.
Narges is an assistant professor in the Departments of Population Health and Radiology conducting research in the Center for Healthcare Innovation and Delivery Science (CHIDS), and a member of its Predictive Analytics Unit.
Her lab's research is focused on the intersection of machine learning, artificial intelligence, and medicine. Using millions of records in the Electronic Health Records database at NYU Langone, as well as hundreds of thousands of imaging and millions of genomic data points, they focus on a number of important topics, including but not limited to: prediction of upcoming preventable conditions and events using machine learning and data science, discovery of disease subtypes using radiology and pathology imaging and electronic records, discovery of existing but undiagnosed medical conditions using electronic health records, and, the discovery of biomarkers and factors associated with important outcomes, etc.