Machine Learning to Optimize Management of Acute Hydrocephalus

Project Summary/Abstract Acute hydrocephalus frequently complicates brain injury including intracerebral (ICH) and subarachnoid hemorrhage (SAH), requiring emergent placement of an external ventricular drain (EVD). The EVD allows rapidly accumulated blood to exit, immediately relieving dangerous increased pressure on the brain. Most patients do not need the EVD after this, while 18-30% develop chronic hydrocephalus and require permanent cerebrospinal fluid (CSF) shunt placement. There is great variability in the management of EVDs across centers, particularly about when to wean EVDs and the approach to surveying and diagnosing EVD-related infection. The longer the EVD is present and the more frequently the EVD is accessed to sample CSF (to test infection), the higher the risk for infection which contributes to high morbidity and mortality. SAH and ICH patients endure EVDs for 11.5-16 days (max > 30), with typical ventriculitis onset occurring at 9.5 days. This vicious cycle is hidden in the cost: 37,000 patients a year receive an EVD in the setting of acute hydrocephalus in the US annually, generating in- hospital charges of $151,672 per patient, or $5.6 billion dollars a year. There is a great need to optimize EVD management by recognizing EVD-related infection while reducing CSF sampling and accurately determining need for permanent shunting (or ability to liberate from temporary drainage), and to do so as early as possible to minimize duration of drainage and length of stay. Our central hypothesis is that there is temporal information in digitized patient data that is reflective of intracranial dynamics that can be harvested to break the negative cycle of ventriculitis and shunt dependency. In previous work, we discovered that intracranial pressure waveform morphology changes two days prior to the clinical diagnosis of ventriculitis. Additionally, we identified a predictor of future CSF shunt dependency as early as four days after EVD placement, building on the correlation of radiographic hydrocephalus changes with concurrent CSF drainage volume. We aim to develop a multicenter purpose-built dataset for the management of acute hydrocephalus including physiologic data such as intracranial pressure waveform, imaging, and clinical data. Using this dataset, we will be able to improve and validate our machine learning models for detection of ventriculitis and prediction of shunt dependence. We will leverage the diversity of the data inputs for model generalizability while also identifying and reducing bias by using a Federated Learning framework for model training and validation. Finally, we will survey physicians to evaluate decision making around EVD management and assess openness to adopting computed prediction scores.