Introduction/Purpose Aneurysmal subarachnoid hemorrhage (aSAH) requires an extended intensive care unit (ICU) stay to monitor for 2 complications, hydrocephalus (HCP) and cerebral vasospasm.This research develops a machine learning (ML) model using quantitative electronic health record data, trended across time, to predict the risk of these 2 complications. Our hypothesis is that the trained ML model can identify patients at high or low risk of cerebral vasospasm and/or shunt dependency. In addition to identifying patients at high risk for complications, the model may also identify patients at low risk for complications, and who consequently might be placed in a ‘fast track’ pathway for expedited transition out of the ICU, reducing ICU-related complications, and allowing more rapid progression to rehabilitation or discharge, thereby reducing hospitalization costs and potentially use of high-cost resources/procedures.

Materials and Methods We have a retrospective 10-year dataset of consecutive patients with aSAH from our tertiary cerebrovascular referral center. A total of 335 demographic, clinical, and radiographic data elements have been extracted from 500 adult patients with angiographically-verified aSAH who survived at least 7-days from admission. These include quantitative metrics trended daily, including hourly vitals (mean arterial pressure (MAP), cerebral perfusion pressure), daily mean serum sodium concentration, hourly external ventricular drain output, and daily transcranial doppler (TCD) blood flow velocities of the cerebral vasculature.

Extracted data will be used to train a ML algorithm to predict overall and real-time risk of cerebral vasospasm and HCP. We will use time series forecasting models, such as autoregressive integrated moving average, incorporating these forecasts into linear and logistic classifiers for complication risk prediction. We will explore how a generative transformer-based model can be used to learn dense patient representations that capture patient status during each clinical time point. This model will use patient representations from aggregate prior time points to predict the likelihood of cerebral vasospasm and HCP events during subsequent time points, by sampling the model-predicted likelihood of these future events. To further increase the prediction confidence of the model, we will use retrieval augmentation to evaluate how similar patient retrieval can be used to enhance model predictive capacity and distinguish between high and low confidence model predictions.

Results Data analysis, algorithm creation, and testing are in process. Preliminary analysis suggests an increased risk of vasospasm with higher Fischer scores, higher Hunt Hess scores, elevated MAP and elevated TCD values. Simple logistic regression models performed on early admission data from days 1 and 2 show improving sensitivity as time from admission increases. These data also show that the feature attribution for TCD measurements is high as it relates to vasospasm risk.

Conclusion The proposed project utilizes prospective and retrospective data to develop a novel ML model to predict complications associated with aSAH, specifically cerebral vasospasm and HCP. The ultimate goal is to use ML in clinical practice for real-time patient risk assessment and optimization of physician decision making and efficient patient care.

Disclosures M. McGrath: None. R. Sen: None. L. Wang: None. R. Meyer: None. V. Shenoy: None. L. Kim: None. M. Levitt: None. L. Sekhar: None.