AbstractMechanical ventilation (MV) is a crucial intervention in the intensive care unit (ICU) for severely ill patients. However, it can potentially contribute to lung damage due to the opening and closing of small airways and alveoli. This study aims to enhance the accuracy of mechanical ventilation prediction using a comprehensive dataset from the Medical Information Mart for Intensive Care (MIMIC-III). The data were extracted with three time frames, 6, 12, and 24 h. Then, 6 h left as a time gap and the ventilation as well as the mortality during the next 48 h. The proposed model consists of two layers: Layer 1 predicts ventilation and mortality in the ICU, while Layer 2 predicts the duration of ventilation. Classification techniques are applied to identify patients in need of ventilators, employing multilayer multitask long short-term memory (LSTM) models. Regression tasks use neural networks (multilayer perception). The optimum feature subset was obtained using particle swarm optimization (PSO). Additionally, this study examines the correlation between ventilation and mortality among patients with and without acute respiratory distress syndrome (ARDS). The findings of this research can enhance health-care outcomes and inform policymakers about resource allocation in overwhelmed health services. The best results were obtained when utilizing the first 24 h for prediction. The proposed MTL model achieved promising performance of 0.944, 0.923, 0.951, and 0.921 for the first task and 0.971, 0.961, 0.963, and 0.970 for the second task for accuracy, precision, recall, score, and AUC, respectively.
Read full abstract