Machine Learning to Optimize Management of Acute Hydrocephalus Patients

每年有37,000名患者在急性脑积水时接受脑室外引流(EVD) 在美国，每名患者的住院费用为151,672美元，或每年56亿美元。有很棒的 神经重症监护病房中优化EVD管理以降低感染率的动机， 准确确定永久分流的需要，并有效地进行，以将持续时间降至最低 排水和停留时间(LOS)。脑室炎的危险因素包括EVD持续时间、脑脊液 (脑脊液)采样频率、脑室内出血(IVH)和插管技术。严重者 IVH和EVDS患者的脑脊液紊乱限制了脑室炎常规脑脊液分析的价值 预测。而脑室炎的诊断并不准确，只有少数人宣布培养阳性，而所有人仍然 要求抗生素治疗和推迟永久分流。这导致了不必要的经验性抗生素 治疗和增加的LOS(30.8天比22.6天)，以及相关的费用(30,335美元以上)和发病率 (例如艰难梭菌感染、出现耐药病理)。确定的过程 永久性的分流依赖在不同机构之间是不同的，特别是在何时决定 开始断奶EVD或预测延迟解决方案。亚急性期的这些决定决定了 LOS和相关的不良事件、暴露于X光检查以及对潜在不必要的承诺 中枢神经系统中的永久性外来物质，然后带来感染和堵塞的终身风险。的确有 没有准确的非侵入性测试(不会进一步引起感染)来诊断脑室炎，也没有 是否有一种及时的方法来预测急性脑积水后是否需要永久分流。为了填满这个 GAP，我们建议开发一个定量模型，从颅内压(ICP)波形分析到 提高脑室炎诊断的精确度，并准确预测是否需要永久分流。在……里面 在之前的工作中，我们能够非常准确地预测谁需要永久分流 使用24小时钳夹试验期间收集的ICP波形分析。然而，复杂的模型只能是 如果它比传统的临床方法更早或更准确地得到诊断，就是合理的。在预赛中 工作中，我们将原始的颅内压波形聚集在一起，发现了一种脑室炎特有的波形模式 在发送诊断培养的前1天出现。我们的中心假设是存在一个时间上的 颅内压波形中反映颅内动态的定量信号可用于优化急性呼吸窘迫综合征 脑积水的处理。影响和意义：基于ICP的非侵入性定量模型 波形分析诊断脑室炎并准确预测是否需要永久分流 减少脑室炎的两个危险因素--EVD的持续时间和脑脊液采样的频率， 同时还减少了LOS、ICU住院期间的相关不良事件和经验性抗生素。

37,000 patients a year receive an external ventricular drain (EVD) in the setting of acute hydrocephalus in the US, generating in-hospital charges of $151,672 per patient, or $5.6 billion dollars a year. There is great motivation in the neurointensive care unit for the optimization of EVD management to reduce infection rates, accurately determine need for permanent shunting, and to do so efficiently in order to minimize duration of drainage and length of stay (LOS). Risk factors for ventriculitis include EVD duration, cerebrospinal fluid (CSF) sampling frequency, presence of intraventricular hemorrhage (IVH), and insertion technique. Severe CSF disturbances in patients with IVH and EVDs limit the value of routine CSF analysis for ventriculitis prediction. And ventriculitis diagnosis is imprecise, with only a minority declaring culture positivity while all still demanding antibiotic treatment and delay of permanent shunt. This leads to unnecessary empiric antibiotic treatment and increased LOS (30.8 vs 22.6 days), with the associated cost ($30,335 more) and morbidity (e.g. Clostridium difficile infection, emergence of drug-resistant pathology). The process of determining permanent shunt dependence is variable between institutions, particularly around the decision of when to begin weaning the EVD or predicting delayed resolution. These decisions in the subacute period determine LOS and associated adverse events, exposure to radiography, and commitment to potentially unnecessary permanent foreign materials in the CNS, which then carry lifelong risks for infection and blockage. There is no accurate noninvasive test (that does not further introduce infection) to diagnose ventriculitis nor is there a timely method to predict need for permanent shunt after acute hydrocephalus. To fill this gap, we propose developing a quantitative model from intracranial pressure (ICP) waveform analysis to increase precision in the diagnosis of ventriculitis and accurately predict need for permanent shunt. In previous work, we were able to predict with good accuracy who would need permanent shunt placement using ICP waveform analysis collected during a 24 hour clamp trial. However, a complex model can only be justified if it achieves a diagnosis earlier or more accurately than traditional clinical methods. In preliminary work, we clustered raw ICP waveforms and found a pattern of waveforms specific for ventriculitis that appears 1 day before diagnostic cultures are sent. Our central hypothesis is that there is a temporal quantitative signal in ICP waveform reflective of intracranial dynamics that can be harvested to optimize acute hydrocephalus management. Impact and Significance: Noninvasive quantitative models based on ICP waveform analysis that diagnose ventriculitis and accurately predict need for permanent shunt would decrease the duration of EVD and the frequency of CSF sampling, two of the risk factors for ventriculitis, while also decreasing LOS, associated adverse events of ICU stay, and empiric antibiotics.