Modeling seizures in patients with focal epilepsy

摘要 在美国，癫痫每年的估计费用为125亿美元，影响约340万人 美国人，一生的风险约为3%。癫痫最常见的形式是局灶性的， 这意味着癫痫发作是由大脑的某个特定区域引起的预测谁将成为一个重大挑战， 尽管有现代技术和研究，但对治疗反应良好，这主要是由于病灶的异质性。 癫痫分析缉获数据的现有统计方法没有适当地解决内部和外部因素， 癫痫发作随时间的个体间差异。 我们建议利用我们对人类癫痫项目（HEP 1和HEP 2）的访问，这是一项观察性研究 450例局灶性癫痫患者纵向追踪癫痫发作，我们的统计和临床专业知识 开发新的动态预测模型，以预测癫痫发作频率随时间的变化。HEP的每日发作数据 显示（1）具有不同癫痫发作轨迹的个体的亚组和（2）癫痫发作的聚集，其中 患者更可能在癫痫发作后经历随后的癫痫发作。动态预测 模型已经成功地用于癫痫以外的其他临床领域，但它们不允许 轨迹的子组。类似地，事件的聚集已经用Hawkes过程处理， 假设一个同质群体的关联模型。相反，我们试图预测 事件和理解协变量对轨迹子组存在的过程的影响。我们假设 考虑到个体轨迹的亚组和癫痫发作的聚集， 精确预测癫痫发作的结果 我们计划通过以下目标开发用于预测癫痫发作事件的新模型：（i）开发一个 贝叶斯非参数模型用于癫痫发作随时间的动态个性化预测。我们将使用 HEP 1数据集预测纵向癫痫发作计数和发生率，允许轨迹和 将使用HEP 2数据评估方法，（ii）开发一种新的Dirichlet过程混合Hawkes过程 用于对重复事件数据进行个性化预测的模型，该模型将允许对事件进行分组和聚类; 并将与现有的方法来处理结块，和（iii）开发一个R包和一个闪亮的 应用程序来实现和说明新的方法。这项工作的预期成果是 具有临床和统计学意义：a）一种闪亮的应用工具，用于获得纵向 基于癫痫发作史、治疗和其他临床相关协变量的癫痫发作轨迹， 患者和临床医生在患者疾病的个性化过程中更早地确定最佳治疗;以及 B）新方法将与其他癫痫类型和其他条件相关，例如在建模中 多发性硬化症的复发

Abstract With an estimated annual cost of $12.5 billion in the United States, epilepsy affects approximately 3.4 million Americans and carries a lifetime risk of around 3%. The most common form of epilepsy is focal in nature, meaning seizures arise from a restricted part of the brain. It remains a significant challenge to predict who will respond well to treatment despite modern technology and research, largely due to the heterogeneity of focal epilepsy. Existing statistical methods to analyze seizure data do not appropriately address the within- and between-individual variation in epileptic seizures over time. We propose to leverage our access to the Human Epilepsy Project (HEP1 and HEP2), an observational study of 450 patients with focal epilepsy that tracked seizures longitudinally, and our statistical and clinical expertise to develop novel dynamic prediction models for seizure frequency over time. The daily seizure data from HEP show (1) subgroups of individuals with different seizure trajectories and (2) clumping of seizures, in which a patient is more likely to experience subsequent seizures following a seizure episode. Dynamic prediction models have been used successfully in other clinical areas besides epilepsy, but they do not allow for subgroups of trajectories. Similarly, clumping of events has been handled with the Hawkes process in association models where a homogenous population is assumed. Instead, we seek to predict occurrence of events and understand covariate effects on processes where subgroups of trajectories exist. We hypothesize that accounting for subgroups of individual trajectories and clumping of seizures will provide more accurate and precise prediction of seizure outcomes. We plan to develop novel models for prediction of seizure events through the following aims: (i) Develop a Bayesian nonparametric models for dynamic personalized prediction of seizures over time. We will use the HEP1 dataset to predict longitudinal seizure count and occurrence that allows for subgroups of trajectories and will evaluate the methods using HEP2 data, (ii) Develop a novel Dirichlet Process Mixture Hawkes process model for personalized prediction of recurrent event data that will allow for subgroups and clumping of events; and will compare to existing approaches to handle clumping, and (iii) Develop an R package and a shiny application to implement and illustrate the novel methods. The expected outcomes of this work are both clinically and statistically significant: a) a shiny application tool to obtain tailored predictions of longitudinal seizure trajectory based on seizure history, treatment and other clinically relevant covariates that will help patients and clinicians identify optimal treatment earlier in the personalized course of the patient’s disease; and b) the new methods will be relevant to other epilepsy types and other conditions, for example in modeling relapses in multiple sclerosis.