Project Summary
Depression is one of the most common disorders of mental health, affecting 7–8% of the population and causing
tremendous disability to af¿icted individuals and economic burden to society. In order to optimize existing treat-
ments and develop improved ones, we need a deeper understanding of the mechanistic basis of this complex
disorder. Previous work in this area has made important progress but has two main limitations. (1) Most studies
have used non-invasive and therefore imprecise measures of brain activity. (2) Black box modeling used to link
neural activity to behavior remain dif¿cult to interpret, and although sometimes successful in describing activity
within certain contexts, may not generalize to new situations, provide mechanistic insight, or ef¿ciently guide
therapeutic interventions.
To overcome these challenges, we combine precise intracranial neural recordings in humans with
a suite of new eXplainable Arti¿cial Intelligence (XAI) approaches. We have assembled a team of exper-
imentalists and computational experts with combined experience suf¿cient for this task. Our unique dataset
comprises two groups of subjects: the Epilepsy Cohort consists of patients with refractory epilepsy undergoing
intracranial seizure monitoring, and the Depression Cohort consists of subjects in an NIH/BRAIN-funded research
trial of deep brain stimulation for treatment-resistant depression (TRD). As a whole, this dataset provides pre-
cise, spatiotemporally resolved human intracranial recording and stimulation data across a wide dynamic
range of depression severity.
Our Aims apply a progressive approach to modeling and manipulating brain-behavior relationships. Aim 1
seeks to identify features of neural activity associated with mood states. It begins with current state-of-the-art
AI models and then uses a “ladder” approach to bridge to models of increasing expressiveness while imposing
mechanistically explainable structure. Whereas Aim 1 focuses on self-reported mood level as the behavioral in-
dex of interest, Aim 2 uses an alternative approach of focusing on measurable neurobiological features inspired
by the Research Domain Criteria (RDoC). These features, such as reward sensitivity, loss aversion, executive at-
tention, etc. are extracted from behavioral task performance using a novel “inverse rational control” XAI approach.
Relating these measures to neural activity patterns provides additional mechanistic and normative understanding
of the neurobiology of depression. Aim 3 uses recurrent neural networks to model the consequences of richly var-
ied patterns of multi-site intracranial stimulation on neural activity. It then employs an innovative “inception loop”
XAI approach to derive stimulation strategies for open- and closed-loop control that can drive the neural system
towards a desired, healthier state. If successful, this project would enhance our understanding of the pathophys-
iology of depression and improve neuromodulatory treatment strategies. It can also be applied to a host of other
neurological and psychiatric disorders, taking an important step towards XAI-guided precision neuroscience.
1