Unsupervised data representation of imaging and genetic data for machine learning applications
Approved Research ID: 59043
Approval date: August 17th 2020
Artificial intelligence algorithms can be trained on biomedical data to predict medical outcomes, such as disease progression or response to treatment. However, biomedical data, like MRI images or genetic information, can be highly complex. Processing highly complex data requires a lot of computational resources and thus poses a challenge to the development of predictive algorithms.
Our research aims to compress biomedical data to make it more manageable to feed into machine learning models. We plan on using a sub family of Deep Learning tools called representation learning for this data compression. Representation learning has been demonstrated to work effectively on speech signal and natural images, and we believe this technique will be well suited for MRI and genetic data. We will then use this compressed data to train models to predict medical outcomes. More specifically, we will train models to predict cognitive decline related to Alzheimer's disease. We expect that this data compression will result in faster development of predictive algorithms as well as improved prediction accuracy. We estimate that this project will take 36 months to complete. Our research will benefit public health by pushing the development of precision medicine.
Our predictive models can be used as tools in Alzheimer's disease clinical trials to improve treatment efficacy by ensuring the most appropriate individuals are selected. We believe our tools will make drug development more efficient so that clinically tested and approved therapeutic interventions for Alzheimer's disease can be rolled out to the public faster.