Predicting trajectories of psychopathology using multimodal neuroimaging and multi-task learning

点击翻译按钮获取中文摘要

PROJECT SUMMARY / ABSTRACT: Most forms of psychopathology have been increasingly recognized as brain disorders that emerge early in development and persist throughout the lifespan. Given the considerable costs of mental illness, it is imperative to develop ways of identifying adolescents who are the most vulnerable, which may lead to more precise and personalized interventions. Here, we propose using a novel predictive framework that may better capture the neurodevelopmental origins of psychopathology, thereby yielding more accurate predictions of psychopathology. Specifically, we plan to develop multi-task neural networks that are trained on spatial maps from three neuroimaging modalities and yield simultaneous predictions of an individual’s age (“brain age”) and psychopathology (“brain pathology”). By integrating predictions of brain age and brain pathology through this novel multi-task framework, we may derive models with improved predictive power, which would also be useful for uncovering the specific biomarkers that underlie each dimension of psychopathology. To investigate these research questions and replicate our findings, we will use multimodal neurodevelopmental data from two of the largest neuroimaging datasets that also contain three longitudinal timepoints – namely the Human Connectome in Development (HCD) and the Adolescent Brain Cognitive Developmental (ABCD) samples. In contrast to using single-task models, we hypothesize that the multi-task predictions of brain age and brain pathology would be better able at detect individual differences at any given point in time (Aim 1) and such predictions would best map onto within-subject changes throughout adolescence (Aim 2). Further, we will use multiple feature importance methods to identify which brain areas and neural properties added the largest predictive power to our most accurate models (Aim 3). This F31 proposal may prove useful in identifying adolescents who are most vulnerable to psychopathology (“personalization”) and accessing risk earlier in the course of development (“precision”). We will also make our deep learning models publicly available so that anyone could use them to yield out-of-sample predictions, which may have wide-spread applications for neuroimaging researchers and pediatric clinicians. Through the pursuit of these research objectives, the applicant will receive essential training in the following areas: 1) deep/machine learning methods, 2) multimodal neuroimaging, 3) advanced psychopathology, 4) conducting rigorous and reproducible research, 5) professional development as the applicant progresses toward a career as an independent, NIH-funded academic researcher. The assembled training team has substantial expertise in each of these subject domains. With their support, the applicant will develop the theoretical, analytical, and professional aptitude needed to foster his research and career ambitions. Altogether, this F31 proposal will be catalyst to help the applicant in his goal of becoming an independent principal investigator that uses multi-model approaches to delineate the most important determents of psychopathology and predict risk on an individual-basis.