Federated Deep Learning to Accelerate Alzheimer's Disease Research

University Of Southern California, Los Angeles CA

Investigators

Abstract

ABSTRACT In this carefully revised application, we launch “Federate AD” - an international alliance to study Alzheimer’s disease (AD) - which combines innovative AI techniques and distributed computation to address urgent challenges in AD research. In 2023, AD and other dementias will cost the nation $345 billion; the disease kills more people than breast cancer and prostate cancer combined. With 6 million people living with AD in the U.S. alone, there is an urgent need to discover factors in the genome and environment that promote or resist dementia. The high cost of collecting neuroimaging and genomic data has led to very large numbers of small sample studies, each with limited power to detect or verify factors that influence disease onset and progression. To address this, we create a distributed artificial intelligence (AI) platform that allows researchers worldwide to compute on Alzheimer’s disease biobanks, uniting expert teams worldwide to discover factors that influence AD onset and progression in worldwide populations. Our Secure Federated Learning architecture computes on AD biobanks in India, Japan, the U.S., and Europe, numbering over 100,000 MRI and PET scans, to (1) diagnose and subtype dementia, using deep learning in large neuroimaging biobanks, (2) infer brain amyloid and tau burden from less invasive biomarkers, such as clinical data and brain MRI and DWI, and (3) predict who will decline clinically from health or mild cognitive impairment to AD. This work will facilitate clinical trial selection and personalized prognosis. Building on our work creating productive, successful worldwide imaging genomics consortia such as ENIGMA and AI4AD, our major innovations include (1) running AI methods on datasets located across the world, (2) enhancing diversity and reducing bias by also learning from diverse non-European datasets, including South Asian and East Asian datasets. Technical innovations include the use of federated deep learning and data harmonization methods that yield site-invariant predictors that generalize better and yield robust predictive models across ancestries. Building on prior DoD funding, we use secure homomorphic encryption to ensure privacy by addressing serious data leakage problems, in which private data can be deduced from AI models. As a whole, this project will produce a social network of AI innovations for AD, and a toolkit for AD researchers to use and build on for applying AI methods in a distributed or centralized setting. The vast datasets from the U.S. ADSP, UK Biobank, and Indian and Japanese repositories have not previously been collaboratively analyzed, and promise insights into common and ancestry-specific predictors of decline. By dynamically learning from worldwide data, new insights will accrue as new data is added. To maximize this work’s impact on the AD field, we carefully integrate our efforts with NIA-funded initiatives in machine learning and phenotypic harmonization, in which we participate. The resulting AI alliance will better integrate worldwide data into AD research, allowing researchers to participate and learn from each other’s data, expertise, and predictive models.