Development of multimodal disease and trait prediction models for personalized medicine
Genetic predisposition and lifestyle account for 70% of health outcomes, but genomic data is rarely used to guide disease prevention and care for common complex diseases. As many factors are involved in the development of common diseases, including age, environmental exposures, physical fitness, and genetic predisposition, the creation of models that incorporate multiple different data types to predict disease development represents an important and active area of research. At Google, we have previously demonstrated that deep neural networks can be applied to multiple types of medical images in order to help identify and predict disease. Additionally, we have shown that using machine learning model outputs as targets of genome-wide association studies can improve discovery of genetic variants associated with disease-relevant traits.
Our objective is to build on previous experience and success with deep learning models and use the uniquely broad and deep data within the UK Biobank to better predict disease progression, identify novel signals in clinical, imaging data and genetic factors that contribute to disease, and develop novel prediction models that incorporate multiple data types to predict disease diagnosis and progression. Results from our investigations will be disseminated to the broad scientific community through public presentations and publications.