Explainable Deep Learning for Neuromodulation Therapies

Y.-J. Chang, Y.-I. Chen, H.-C. Yeh, S.R. Santacruz
University of Texas at Austin,
United States

Keywords: explainable deep learning, multi-scale modeling, neuromodulation


Neuropsychiatric disorders are the most prevalent and costly illness worldwide. However, standard psychotherapeutic interventions have limited capability of reducing the severity of mental illness. Additionally, uncertainty of side effects due to simultaneous polydrug use and low efficacy resulting from blood-brain barrier make pharmacological options undesirable. As an alternative treatment, neurostimulation is designed to modulate brain activity through the targeted delivery of electrical stimulation. Invasive neuromodulation such as deep brain stimulation (DBS), allowing high spatial resolution for localizing treatment, has proven successful in treating many movement disorders, but has yet to translate well to psychiatric conditions. The key limiting factor is that the true mechanisms of DBS remain unclear, resulting in the failure of developing precision treatments to address the varied outcome due to patent heterogeneity. As neural dynamics illuminate the causal interactions between neurons or networks, it is necessary to develop a neural dynamics-based framework to assess the brain computations governing cognitive functions. To date, the investigation of the neural population dynamics is mostly limited by the single-scale analysis. Although in neuropsychiatric conditions, neural circuit-wide pathological activity impacts dynamics at multiple scales, either directly or indirectly, there is no broadly accepted multi-scale dynamical model for the collective activity of neuronal populations. Traditionally, the analysis has largely proceeded without formal neurobiological models of the underlying multi-scale neuronal activity. Whereas cross-correlation or coherence have been employed to measure the coupling of activities at different scales (e.g., spiking and local field potential (LFP) for neuronal synchronization), they only capture patterns of statistical dependence. Instead, dynamical modeling, which is seldom explored at multi-scale level, infers the causal interactions among brain regions or sources and potentially yields mechanistic understanding of brain computations. Here we developed a neurobiological model-driven deep-learning model, Multi-Scale NeuroBondGraph network (MS-NBG), to uncover multi-scale brain communications governing cognitive behaviors. Our previous work has demonstrated the success of characterizing cross-scale field potential interactions in center-out joystick task. By adding a gated recurrent unit (GRU) as an initial state estimator and implementing the GRU-based recurrent structure, MS-NBG further improves the performance and the efficiency. We implemented MS-NBG on spike trains and LFPs, which are commonly studied neural signals, from macaques to study communications in the brain networks. The MS-NBG accurately reconstructs spikes and LFPs with a root mean squared error of 2.47 (μV) for LFP and 0.22 (s-1) for spiking. Multi-scale effective connectivity extracted from the MS-NBG illustrates the influence that a source exerts over another signal at multiple levels. Graph theoretical analysis is employed to characterize the network’s integration and segregation, which suggests explanations of naturalistic behaviors. With the incorporation of realistic brain constraints in the MS-NBG, this neurobiologically realistic framework holds great potential to improve our understanding of complex and causal brain functions in unprecedented detail. With the application of stimulation technique such as DBS, we are able to characterize the variation of large-scale brain dynamics and identify the “neural-marker” that can interpret the alteration of neural activity and underlying behavioral variables to design the optimal and precise treatment.