Crowd Sourced Development and Validation of Neuro-Computational Models of Affective Processes*
Crowd sourced development and validation of neuro-computational models of affective processes. Objective biomarkers of pathology exist for a number of diseases, and their development is one of the great advances of modern allopathic medicine. However, objective assessment of affective processes related to mental health disorders has lagged far behind. Currently, the only way to diagnose mental illnesses like depression is using self-reported symptoms such as increased feelings of sadness, guilt, or irritability and decreased interest in activities, concentration, and energy. Yet, these symptom-based diagnoses are astonishingly unreliable across providers (Kappa coefficient = 0.25)1 likely due to how such illnesses inherently degrade individuals’ ability to accurately make these judgments (e.g. depressive realism bias 2). Thus, developing reliable objective biomarkers could dramatically improve diagnosis and treatment by allowing mental health to be characterized on the basis of underlying neuropathology rather than external self-reported symptoms. Direct measures of brain function provide a promising area for developing biomarkers of emotion pathophysiology. In the past several years, major advances in combining functional magnetic resonance imaging (fMRI) with machine learning techniques—algorithms for finding predictive patterns in complex datasets—have brought the goal of fMRI-based assessment of affect within reach. We have recently demonstrated for the first time that fMRI activity can predict whether an individual person is experiencing high or low emotional responses to arousing pictures with over 90% accuracy 3. Critically, this biomarker is sensitive and specific to emotional responses, when compared with other salient and arousing affective events such as thermal pain.
This preliminary success raises a number of issues that must be addressed before fMRI-based biomarkers can be used in large-scale clinical trials and clinical practice, including demonstrating: a) robustness across laboratories and procedures, b) specificity to type of emotion and elicitation method, c) applicability to clinical populations, and d) sensitivity to responses in clinical interventions.
Neurolearn will leverage existing code from Neurosynth.org 4 and Neurovault.org 5 for uploading and storing image data, interactively viewing brain images, and searching and selecting data. Novel features that we will develop in our prototype include: (a) a database for neuroimaging “person-level maps” and “research products” (i.e., brain signature) data types; (b) meta-data specification and entry; (c) user accounts and sharing permissions; (c) server-side machine learning algorithms to generate new signature maps; and (d) server-side applications to apply signature maps to subject data and evaluate sensitivity/specificity related to user-specified outcomes. Finally, we will deploy our application to a local webserver that will be optimized to accommodate both the large storage demands for the imaging repository and the high RAM and CPU cores needed to simultaneously process multiple queued user jobs.