Predicting Hospital Length of Stay Using Residual Neural Networks with Self-Attention: A Deep Learning Approach

To effectively manage the resources of a hospital, it is crucial to know the length of stay (LOS) of a patient in the hospital. With this problem in mind this paper proposes a progressive deep learning model that combines Residual Neural Networks (ResNets) with self-attention mechanisms to predict LOS as a regression problem. The proposed model leverages the power of residual connections to facilitate efficient learning in deep architectures and incorporates self-attention to focus on the most relevant feature interactions, capturing both local and global relationships in the data. The dataset used consists of comprehensive patient demographics, hospital attributes, and admission details. After training for 100 epochs, the model achieved a mean absolute error (MAE) of 2.35 and a root mean squared error (RMSE) of 3.22 on the training set, and a MAE of 2.55 and RMSE of 3.75 on the validation set. These results demonstrate the efficacy of the ResNet with self-attention in predicting hospital LOS with high accuracy, outperforming traditional methods. The findings highlight the potential of using advanced deep learning models to enhance hospital management systems and improve healthcare delivery by effectively predicting patient LOS. Further research can build on this model to include additional clinical factors and optimize healthcare resource utilization.