Predicting Decisions and Disorders from Brain Images

Abstract

Researchers have long been fascinated with the concept of reading human minds by studying their brain physiology. Advancements in neuroimaging technology in conjunction with the application of computer vision and machine learning to analyze brain images have made mind-reading a realistic goal. Large-scale ventures are currently being pursued in both academic and industry research to develop interfaces that allow direct communication with machines by analyzing brain signals. As such, efficient methodologies to model neuroimaging data to infer mental states are highly warranted. Modeling neuroimaging data is a challenging task since it is recorded over time, noisy, multi-dimensional, and massive in size. Recent availability of large volumes of neuroimaging data has facilitated the use of deep learning which is shown to outperform classical machine learning methods at modeling and inference.In this work, we developed and compared several deep-learning methods to efficiently model neuroimaging data to predict choices and detect early stages of neurological disorders. First, we adapted concepts from natural language processing on functional neuroimaging data to predict choices involving everyday experiences and money. Second, through the application of advanced deep learning methods on multi-modal neuroimaging data obtained during dual tasking, we identified mild cognitively impaired individuals from the normal population. Finally, we developed a deep-learning framework to predict choices and detect decision-making impairment using neuroimaging data extracted from the participants engaged in the Iowa Gambling Task. This dissertation presents efficient ways to model neuroimaging data collected under different conditions. Our results support theoretical notions behind decision-making. We establish suitable baselines and validation methods to evaluate the performance of deep-learning models in predicting decisions and disorders using brain images.