Abstract 405: A Machine Learning-Based Prediction of Cardiac Arrest Outcome Using a Large Multi-Center Database

Abstract

Background: Cardiac arrest (CA) poses a significant risk of death and long-term neurological disability yet there remains a lack of accurate and reliable methods to prognosticate post-CA outcome, indicating a major unmet need. Additionally, no existing CA outcome prediction model utilizes physiological time series (PTS) data, which captures transitions in a patient’s health state at high temporal resolution. Materials and Methods: We screened the Philips eICU database, which includes over 200,000 patients in 200 hospitals, extracting data on CA patients within 24 hours of ICU admission. We first extracted PTS and Electronic Health Record (HER) features and selected our features using nested Random Forest. The outcome labels were neurological state or mortality at ICU discharge. We tested various individual and ensemble machine learning algorithms (GLM, Random Forest, Gradient Boosting, Support Vector Machine, Neural Networks) and optimization techniques (hyperparameter tuning, model stacking, and transfer learning) to maximize the accuracy of outcome prediction. Results: For prediction of neurological outcome, our model combining EHR and PTS features achieved a higher sensitivity (0.78), specificity (0.88), and AUC (0.87±0.01) compared to the APACHE baseline standard model (AUC: 0.74±0.01, sensitivity: 0.77, specificity: 0.63). Also, for prediction of mortality our model (AUC: 0.81±0.01, sensitivity: 0.78, specificity: 0.71) outperformed the APACHE clinical baseline (AUC: 0.75±0.01, sensitivity: 0.86, specificity: 0.56) by 7%. The integration of PTS with EHR data increased mean AUC by 4% for neurological outcome and 2% for mortality. Conclusion: Results demonstrate that for both neurological outcome and mortality prediction, our ensemble model was significantly more effective compared to the baseline APACHE model. Additionally, we find that PTS data adds to the discriminative power of clinical models. With appropriate validation in prospective cohorts, these findings could be used as point-of-care references to aid ICU physicians in clinical decision-making.