Random Forest and Attention-Based Networks in Quantifying Neurological Recovery

Introduction: Cardiovascular disease is generally considered the most prevalent cause of morbidity in the modern world, and cardiac arrest, in particular, causes nearly 50 % of deaths linked with heart attack and stroke in the US. Surviving cardiac arrest could still lead to brain injury and, consequently, death. Our main aim is to mitigate incorrect prognoses in measuring patients' recovery by exploiting the power of machine learning. Methods: We use the training set from the unofficial phase comprising 607 comatose adults following recovery from cardiac arrest to develop two attention-based networks using various features. 486 subjects are used for training and 10-fold cross-validation; the remainder is used for testing and evaluation. Results: Despite an official challenge score of 0.00, Team_KU's best attention-based models yielded a testing accuracy of 62.00 %, an F-measure of 61.20 %; beating our random forest used in the unofficial phase at 55.58 %, and an area under the receiver operating characteristics (AUC) of 0.63 for outcome classification and a mean absolute error of 2.49 for CPC prediction with 607 subjects; nearly half of the provided data in the official phase. Conclusion: This study paves the way toward implementing efficient machine learning to assess brain injury in comatose patients, even in resource-restricted settings. Thus allowing early, automated prediction of recovery.