Background: In-hospital cardiac arrest (IHCA) remains a leading cause of mortality in the United States, with an estimated 290,000 cases annually and survival to discharge rates ranging from 15% to 40% as of 2024. For those patients who achieve return of spontaneous circulation (ROSC), predicting their neurological outcome and survival in the intensive care unit is critical for guiding treatment decisions and prognostication. The Glasgow Coma Score (GCS) provides a standardized measure of consciousness corroborated by salient putative measures of neurologic function. This study investigated the importance of incorporating GCS, assessed independently of pharmacologic sedation, into a machine learning model designed to reveal variance explained (VE%) and by utilizing routinely available data to predict hospital mortality in post-ROSC patients surviving at least one day of intensive care.
Methods: Patient demographics, ICU Day 1 laboratory results, and administrative data were extracted under IRB exemption from the EMR. Generalized regression with LASSO and response surface methodology identified from putative and biologically plausible features those associated (p<.05) with hospital mortality. Bootstrap Forest ensemble modeling estimated each feature’s proportion (%) of variance explained (VE) in the overall model’s accuracy (R2) to estimate mortality risk. Continuous data summarized with median [IQR] were compared using the Kruskal–Wallis test. Discrete data summarized as proportions were compared with the chi-square test. A two-tailed p< 0.05 with confounders controlled including concurrent SARS-CoV-2 infection was considered significant.
Results: Demographics, comorbidities, and clinical features independently associated with hospital mortality are presented in Table 1. Explained variance in clinical features predicting mortality with versus without sedation evoked modification of GCS are provided in Table 2. While GCS remains a valuable prognostic indicator, findings reveal a dynamic interplay between sedation and other physiological factors. Notwithstanding the isomorphic accuracy of modeling (R2), sedation blunts GCS predictive power (11% VE) compared to when patients are not sedated (28% VE).
Conclusions: When patients are sedated, the GCS loses some predictive power emphasizing the need for nuanced interpretation of complex interplay among various factors influencing mortality in post-ROSC patients. Depth of sedation likely plays a role in how much it influences GCS and other predictors. Machine learning facilitated an understanding of how sedation impacts these factors and can improve the ability to provide personalized and effective care. In sedated patients, kidney function (eGFR, creatinine) and urine output emerged as stronger predictors of mortality, suggesting a complex relationship between sedation, neurological status, and overall systemic health. Recognizing the amplified importance of factors such as kidney function and urine output in sedated patients suggests closer monitoring of fluid balance and renal support may be warranted.
