Mixed graphical models for the prediction of neurological morbidity in the PICU

在现代儿科重症监护病房（PICU），随着死亡率的持续下降，重点已转向降低神经系统发病率的措施。由于病情严重，儿童经常接受镇静和/或神经肌肉阻滞治疗，因此在PICU中很难发现神经系统并发症。早期识别和实施循证治疗策略对减少神经系统发病率至关重要。传统的神经监测方法（计算机断层扫描[CT]、磁共振成像[MRI]、脑电图[EEG]）不能实际用于常规筛查目的。我们认为，结合生物标志物和临床数据的生物数学模型可能是检测PICU神经系统并发症的重要工具。这一策略可能允许快速识别神经系统并发症和早期干预，最终降低发病率和死亡率。在这个指导的以患者为导向的研究职业发展奖项中，我们将尝试使用由共同赞助者开发的新算法MGM-Learn（混合图形模型学习）来开发混合图形模型，该算法具有处理连续和离散变量的独特能力。匹兹堡大学匹兹堡儿童医院PICU收治的228名诊断不同的儿童将被纳入。血清生物标志物（髓鞘碱性蛋白[MBP]、S100B、脑源性神经营养因子[BDNF]和胶质纤维酸性蛋白[GFAP]）在预测神经损伤（如创伤性脑损伤或心脏停搏）后的预后方面显示出前景，将与从电子健康记录中获得的临床和实验室变量结合使用。通过混合图形模型的综合分析，预测入院时不存在的神经系统并发症的急性发展（如癫痫、中风、出血、脑病）和出院时和危重疾病后6个月的发病率（如功能状态量表（FSS）、儿科生活质量量表（PedsQL））。这次K23奖将为我提供混合图形建模方面的深入培训，极大地提高我在神经生物标志物临床应用方面的技能和有效的领导和管理能力，使我成为一名成功的独立研究者。它将为我的R01提供初步数据，早期预警神经生物传感器系统的实施，通过使用混合图形模型，不断填充最新的生物标志物和临床数据变量，进入临床实践，实时检测神经系统并发症；并评估其通过早期识别神经系统并发症和及时执行治疗策略以防止不可逆脑损伤来降低神经系统发病率的能力。

In the modern pediatric intensive care unit (PICU), as mortality rates continue to decline, focus has shifted towards measures to decrease neurological morbidity. Neurological complications can be difficult to detect in the PICU as children oftentimes receive sedation and/or neuromuscular blockade due to the severity of their illnesses. Early identification and implementation of evidence-based treatment strategies is paramount to the reduction of neurological morbidity. Traditional methods of neuro-monitoring (computed tomography [CT], magnetic resonance imaging [MRI], electroencephalography [EEG]) cannot be practically utilized for routine screening purposes. We believe that biomathematical models integrating biomarkers and clinical data may represent an important tool for the detection of neurological complications in the PICU. This strategy may allow for rapid identification of neurologic complications and earlier intervention to ultimately reduce morbidity and mortality. In this mentored patient-oriented research career development award we will attempt to develop mixed graphical models using a novel algorithm developed by the co-sponsor, MGM-Learn (Mixed Graphical Model Learning), which has the unique capability of processing continuous and discrete variables. Two hundred and twenty-eight diagnostically diverse children admitted to the PICU at Children's Hospital of Pittsburgh of UPMC will be enrolled. Serum biomarkers (myelin basic protein [MBP], S100B, brain derived neurotrophic factor [BDNF], and glial fibrillary acidic protein [GFAP]) that have shown promise in prognostication of outcome after neurological injuries such as traumatic brain injury or cardiac arrest will be used in conjunction with clinical and laboratory variables obtained from the electronic health record, through integrative analysis in mixed graphical models to predict acute development of neurological complications that were not present at the time of admission (e.g. seizure, stroke, hemorrhage, encephalopathy) and morbidity (e.g. Functional Status Scale (FSS), Pediatric Quality of Life Inventory (PedsQL)) at discharge and 6 months following critical illness. This K23 award will provide me with in-depth training in mixed graphical modeling, greatly enhance my skills in the clinical application of neuro-biomarkers and effective leadership and management to transition to a successful independent investigator. It will provide preliminary data for my R01, the implementation of an early warning neuro-biosensor system, through the use of mixed graphical models that continually populates with the most up-to-date biomarker and clinical data variables, into clinical practice to detect neurological complications at a moment to moment basis; and the assessment of its ability to reduce neurologic morbidity through early recognition of neurological complications and timely execution of treatment strategies to prevent irreversible brain damage.