(731) Inpatient Mortality Prediction Model in Critically Ill Intracerebral Hemorrhage Patients

Introduction: Intracerebral hemorrhage (ICH) is a common stroke type associated with high morbidity & mortality. Its significant global impact has led to efforts in developing outcome prediction models for improved prognostic certainty and early treatment identification. However, existing models primarily emphasize neurological factors and lack validation in diverse, large populations. Using the eICU Research Database and Medical Information Mart for Intensive Care IV [MIMIC] datasets, we studied various physiological and clinical factors and its association with mortality in ICH patients, and developed an inpatient mortality prediction model based on selected variables.

Methods: Using the eICU database, adult patients with non-traumatic ICH were selected after excluding incomplete data (n=1166). Variables used for development of prediction model were identified clinically and included age; APACHE III score; Glasgow Coma Scale [classified into severe (GCS < 8), moderate (8 ≤ GCS < 12), mild (12 ≤ GCS ≤ 15)]; lab values upon ICU admission (INR, creatinine , hemoglobin, platelet count, PTT ); temperature and mean arterial pressure "burden" in the first 72 hours of ICU admission [the area under the curve of temperature/blood pressure time series for thresholds febrile (>38°C), hypertension (MAP>85 mmHg), and hypotension (MAP < 55 mmHg)]. A set of models of inpatient mortality was then developed and validated using patient data from the MIMIC IV dataset with the same exclusions and input variables (n=1078).

Results: In a 5-fold cross validation within eICU, the top three models developed are logistic regression [LR] (accuracy: 0.8482 (95% CI: 0.7821, 0.9143)), a random forest classifier [RFC] (accuracy: 0.8439 (95% CI: 0.794, 0.8938)), and a multi-layered perceptron [MLR] (accuracy: 0.8293 (95% CI: 0.7826, 0.876)). External validation was done on MIMIC IV, which showed LR (accuracy: 0.8135 (95% CI: 0.7892, 0.8357)), RFC (accuracy:0.8089 (95% CI: 0.7844, 0.8313)), and MLP (accuracy: 0.8015 (95% CI: 0.7766, 0.8242)).

Conclusions: A supervised machine learning model based on systemic physiologic and clinical variables is able to accurately predict mortality in ICH patients across different patient datasets. Next steps are to create a model that can identify early treatment targets to reduce mortality.