Interpretable, subject specific-mapping of neurological health in the performance setting

Approved for Public Release: Next generation objective tools are needed to guide personnel selection both upon entering the Navy and over a career. Such an approach could serve as a powerful tool in monitoring performance, determining operational readiness, and effective placement of personnel. While sophisticated tools exist for acquiring information on brain structure and function, there persists a critical need for the development of interpretable constructs for integrating this information into key, actionable metrics. We propose development of Normative Neurological Health (NNH), a data-driven panel for assessing specific brain features in military populations. This #brain basis set# will provide overall scores but will also parse neurological features into both neurobiological and behavioral components. These can be leveraged during recruitment, referred to over time and monitored in relationship to training, field experience, and performance. The NNH will be optimized for interpretability by decision makers, to maximize the assignment and longevity of talent and to mitigate risk as personnel progress through their career. The primary goal for the NNH is a practical tool that can be used in the human performance setting, enabling real world evaluation. The NNH will be implemented by combininglarge normative datasets with machine learning and will be designed, from the start, with active learning in mind i.e. the ability to update both the reference normative data and the model as additional data accrues.For the current effort, we will implement a framework utilizing a battery of derived neuroimaging measures to objectively identify levels of resilience and sensitivity in cognitively normal individuals. The utility of the NNH framework will arise from learning meaningful and objective embeddings of brain data into a latent space that represents individual-level subtypes. Subtypes are smaller, distinctive groups within an overall population. Subtyping must be predictive and follow a machine learning style with a #training# phase for the models followed by inference wherein we apply the fixed models to new data. The embedding methodology, deep SiMLR, will build upon the SiMLR algorithm developed under a recently completed ONR award (ONR #N00014-18-1-2440). SiMLR has already received national and international recognition as a groundbreaking tool for performing true multi-modal analyses of advanced imaging and non-imaging data within a joint analytical framework. The next generation version of SiMLR will be implemented with deep learning and thereby immediately gain representation power (due to the expressive power of deep architectures), improved computational efficiency, access to more flexible loss functions and adversarial optimization strategies. In order to train deep SiMLR to learn meaningful embeddings that can be mapped to spaces that are relevant to sailor performance, we will also leverage large training data resources, including UK Biobank (UKBB) and the Human Connectome Project (HCP). We will identify NNH metrics via deep SiMLR that are objective, rapidly computable and interpretable in terms of performance. The NNH will be highly relevant to monitoring personnel over a career. We will use the best available assemblage of data in order to establish the relevance of the NNH. An extended evaluation aim will take advantage of the variance between existing #super-normal# populations and populations that include subjects with known risk exposures # thereby demonstrating efficacy in termsof achieving both (a) the expected performance-related ranking of brain data and (b) further generalizability of the NNH to more diverse data. Applied in a targeted manner, the NNH will improve the effectiveness and efficiency of naval operations and provide a new window into the dynamics of neurological capability and performance in naval personnel.