Invasive decoding and stimulation of altered reward computations in depression

摘要 抑郁症是一种非常普遍的心理健康疾病，影响着美国数百万人， 对福祉、残疾率和医疗保健费用的影响。尽管有这些巨大的影响和成本， 目前用于治疗抑郁症的治疗选择的有效性是有限的。近年来， 在非侵入性成像研究结果的指导下，努力开发脑深部电刺激（DBS）策略。 不幸的是，这些药物未能显示出显著的疗效，可能是因为疾病的巨大异质性。 这是因为缺乏足够的数据来了解疾病的神经生物学原因。当前方法 使用非特异性生物（药理学）或解剖（DBS）靶点，导致部分有效性和副作用。 对抑郁症基础的患者特异性诊断将允许量身定制的治疗设计， 在研究神经功能的标准方法中，最近的方法利用多区域 神经外科癫痫患者中的侵入性电生理记录，这些患者通常患有共病抑郁症， 收集高质量（多区域、高信噪比、高时间分辨率）神经生理学数据，并具有 允许患者特定疾病模型和靶向神经刺激的潜力。此外，机器学习 方法允许将这种高维神经数据映射到受抑郁症集中影响的患者的情绪状态， 并强调了边缘区域的单点和跨区域活动的参与，包括海马体，杏仁核， 和眶额皮质然而，解码方法本身存在若干挑战。首先，他们选择特征 与自我报告的情绪有关-一个复杂的，抽象的概念，缺乏客观，定量的基础， 可能在不同的患者中报告不同，使得泛化具有挑战性。其次，数据驱动的方法不是 基于目前的大脑功能模型，因此缺乏对疾病基础的机械解释。第三、 情绪是在没有明显行为的情况下报告的，这使得难以框定行为功能的缺陷。 受影响的大脑区域。最后，目前的方法缺乏解码的神经特征和 激励策略。在这里，我们建议通过结合分布式iEEG记录， 决策的强化学习模型和机器学习方法来研究奖励处理 相关脑区的癫痫患者和没有共病抑郁症。我们将检查本地激活， 以及决策过程中眶额皮层、杏仁核和海马体的电路动力学（功能连接）， 做出行为。我们将寻求在决策的强化学习模型中建立神经生理学数据， 提供影响情绪的定量、可重复的行为指标，并建立在长期观察的基础上 关于抑郁症中受损的奖励处理。最后，我们将开发患者定制的刺激范例， 抑郁症受我们对患者特异性疾病表现的观察限制。我们预计， 带有奖励模型的侵入性记录将为理论上的计算表征打开大门 抑郁症的神经行为缺陷，并允许开发新的治疗策略。

Abstract Depression is a highly prevalent mental health disorder that affects millions of people in the US and causes significant impacts on well-being, rates of disability and health care costs. Despite these substantial impacts and costs, the effectiveness of current therapeutical options for the treatment of depression is limited. In recent years there have been efforts to develop deep-brain stimulation (DBS) strategies guided by results from non-invasive imaging studies. Unfortunately, these have failed to show significant efficacy, likely because of the vast heterogeneity in disease presentation and the lack of sufficient data to understand the neurobiological causes of the disease. Current approaches use unspecific biological (pharmacology) or anatomical (DBS) targets, leading to partial effectiveness and side effects. Patient-specific depictions of the basis of depression would allow tailored treatment designs, but data of sufficient quality is mostly unavailable with standard approaches to the study of neural function. Recent approaches leverage multi-areal invasive electrophysiological recordings in neurosurgical epilepsy patients, which often suffer from co-morbid depression, to collect of high-quality (multi-areal, high signal-to-noise, high temporal resolution) neurophysiological data, and have the potential to allow patient-specific models of disease and targeted neurostimulation. In addition, machine learning methods allow mapping this high-dimensional neural data onto patient’s emotional states centrally affected in depression, and highlight the involvement of single-site and cross-areal activity in limbic regions, including the hippocampus, amygdala and orbitofrontal cortex. However, decoding methods present several challenges of their own. First, they select features associated with self-reported mood - a complex, abstract concept that lacks an objective, quantitative foundation and may be reported differently across patients, making generalization challenging. Second, data-driven methods are not grounded on current models of brain function, and thus lack mechanistic explanations of disease underpinnings. Third, mood is reported in the absence of overt behavior, making it difficult to frame the deficits in the behavioral functions subserved by affected brain areas. Finally, current approaches lack a connection between decoded neural features and stimulation strategies. Here, we propose to address these challenges by combining distributed iEEG recordings, reinforcement learning models of decision-making and machine-learning approaches to study reward processing in relevant brain areas from epilepsy patients with and without comorbid depression. We will examine local activations as well as circuit dynamics (functional connectivity) in orbitofrontal cortex, amygdala and hippocampus during decision- making behavior. We will seek to ground neurophysiological data in reinforcement learning models of decision-making to provide quantitative, reproducible behavioral metrics that impact mood, and to build on long-standing observations regarding damaged reward processing in depression. Finally, we will develop patient-tailored stimulation paradigms for depression constrained by our observations on patient-specific disease presentation. We expect that the combination of invasive recordings with reward modeling will open the door to a theoretically-grounded computational characterization of neurobehavioral deficits in depression, and allow development of novel treatment strategies.