A deep learning approach for length of stay prediction in clinical settings from medical records

Deep neural networks are becoming an increasingly popularsolution for predictive modeling using electronic health records because of their capability of learning complex patterns and behaviors from large volumes of patient records. In this paper, we have applied an autoencoded deep neural network algorithm aimed at identifying short(0-7 days) and long stays (> 7 days) in hospital based on patient admission records, demographics, diagnosis codes and chart events. We validated our approach using the de-identified MIMIC-III dataset. This proposed Autoencoder+DNN model shows that the two classes are separable with 73.2% accuracy based upon ICD-9 and demographics features. Once vital chart events data such as body temperature, blood pressure, heart rate information available after 24 hour of admission is added to the model, the classification accuracy is increased up to 77.7%. Our results showed a better performance when compared to a baseline random forest model.