Abstract:
Treatment notes from the pediatric ICU can be effectively used to construct a predictive model for the onset of shock with a lead time of 24 hours. If this disease is not diagnosed at an early stage, it can create multiple complications that might also lead to the patient’s death. In this project, we try to build state-of-the-art algorithms in natural language processing such as word2vec to build shock prediction models. This might help the doctors take care of severe and sensitive patients at more regular intervals to prevent such complexities. We train machine learning models using features we extract from word embeddings based on novel techniques. The evaluation of the models on the test data is based on accuracy, precision, recall, F1-Score and AUC. We demonstrate clinical interpretability of these models, thus making these widely accessible to critical care physicians.