Machine learning algorithm to predict in-hospital mortality after aneurysmal subarachnoid hemorrhage

Machine learning (ML) methodology surpasses the traditional tools of statistical analysis in processing big data clinical datasets . Aim. To develop an ML algorithm of application of recurrent neural network to analyze clinical datasets of patients with aneurysmal subarachnoid hemorrhage (SAH). Materials and methods. A big data registry included retrospective data from 2,631 patients with an arterial aneurysm. From these, 390 individuals were selected who required treatment for SAH in an intensive care unit (ICU) setting. The raw dataset contained 7290 features, from which 12 features were selected to train the following ML models: logistic regression, support vector machine, random forest, XGBoost, multilayer perceptron and long short-term memory network (LSTM) were tested. Data preprocessing and modeling were provided in Python (version 3.11.4) using scikitlearn, tensorfl ow, keras and hyperopt libraries. The values and 95% confi dence intervals (CI) of AUROC and AURPC, predictive value, specifi city and sensitivity were calculated. Results. We recruited 246 (63%) females and 144 (37%) males with mean age of 54±12.9 years. Death occurred in 133 (34%) patients including 33 patients deceased during 24 hours after admission. The best model for predicting lethal outcome was LSTM. After comparison with other ML algorithms LSTM showed the highest predictive values (AUROC – 0.83; 95% CI: 0.72–0.92, AURPC – 0.62; 95% CI 0.39–0.81) in term of in-hospital mortality. For the period in ICU from day 3 to day 6, the model’s positive predictive value was 0.83, sensitivity 0.95 and specifi city 0.58. Conclusions. LSTM may be applied to development of automatic algorithms in management of critically ill patients after SAH.