Real-time Detection of Delayed Cerebral Ischemia using Machine Learning

In the field of neurological intensive care (NICU), the duration of ischemic burden (or inadequate blood flow) on neurons is known as a primary way in which functional outcomes can be affected. There is a therapeutic time window when interventions may prevent reversible ischemia from becoming permanent infarction in stroke. Decreasing the process time from onset of stroke symptoms to intervention has shown benefit for patient outcome. Delayed cerebral ischemia (DCI) is the leading cause of morbidity after subarachnoid hemorrhage (SAH), and occurs in 30-54% of patients while in the NICU. Yet it is the only type of intervenable cerebrovascular insult for which there is no method to identify time of onset. Without knowledge of the therapeutic window, interventions to reverse DCI are biased to fail both in a research setting (without accurate timing of events, retrospective data analysis is uncertain) as well as at the point of care (by the time DCI is recognized, the therapeutic window may have passed). Despite the high impact on patient outcomes, there are many barriers to timely detection of DCI. Symptom onset can be subtle and develop gradually, and clinical diagnosis often relies on a change in exam, though 20% of SAH patients are comatose. Physicians must surpass a certain threshold of suspicion to warrant the health risks of confirmatory testing (radiation exposure, laying patient flat), which includes ruling out clinical mimics. There is an opportunity for new technology to help improve the timely diagnosis of DCI. Our hypothesis is that physiologic data and in particular their temporality are amenable to developing a real-time detection tool for DCI. To fill the scientific and clinical gap, we propose using recurrent neural networks to build and refine a sequential learning model that detects DCI onset and distinguishes clinical mimics, and then translate it into a clinical decision support tool. SA1: Build a recurrent neural network model for real-time detection of DCI. SA2: Implement a pipeline to deliver the Clinical Decision Support system. A sequential model to detect DCI onset and differentiate clinical mimics would fill a scientific gap, providing a therapeutic window. A novel clinical decision support tool to deliver this information, implemented in a hospital setting, would allow individualized approaches to urgent testing of clinical mimics and enable critical timely action to reverse ischemia and reduce morbidity (AHA Program: Postdoctoral Fellowship)