In the last decade, Deep Neural Network models (DNNs) have facilitated remarkable advances in brain decoding, revealing how distinct patterns of neural activity correspond to different thoughts, sensations, and behaviors. From brain activity alone, these models can predict what images a person is seeing or imagining in each moment, current states of pain, specific auditory sensations, and many other mental states. When trained on sufficiently large data, they can not only be used to decode current mental states, but also predict specific phenotypes (e.g., average weekly alcohol intake, fluid intelligence, etc.). In this investigation, we will combine DNNs with a “big data” approach to decode cognitive and behavioral phenotypes from existing brain data samples collected at the University of Florida. Specifically, by leveraging the computing power of the HiPerGator supercomputer, we will build predictive models from the large, openly accessible UKBioBank dataset, and then apply these models to predict cognitive and behavioral phenotypes from specific populations of interest (e.g., aging individuals). This work will (i) facilitate development of (and make publicly available) a new research algorithm for deep learning, and (ii) provide insights into how brain structure and function are linked to different phenotypes. The duration of this project is from January 1, 2021, through January 1, 2022. The public health impact of this project is that by using deep learning, we will be able to (i) identify predictive markers for maladaptive phenotypes (e.g., alcohol abuse) that can be used to inform behavioral and therapeutic interventions, and (ii) identify brain regions of interest for future targeted neurofeedback interventions with fMRI, such as using neurofeedback to decrease cigarette, food, and alcohol cravings.
