A Few-Shot Learning-Based Framework for Predicting Likelihood of In-patient's Mortality.

In recent years, with the evolution of technology and electronic health records (EHR), the digitization of medical information enables the collection and analysis of vast amounts of data, encompassing physiological records, disease history, and medication records. This research aims to enhance healthcare delivery based on EHR, particularly predicting the likelihood of in-patient's mortality and length of stay. However, this digitization also brings challenges in data analysis, including managing high dimensionality, sparsity, imbalance, and unstructured features of data. Furthermore, the model's generality and the handling method of indicators with high correlation must be considered in the research. An efficient approach is required to deal with situations where limited data or task-related scenarios could benefit from information sharing. Despite the challenges posed by EHR's high dimensionality, sparsity, imbalance, and unstructured features, digital health technologies offer significant potential for enhancing medical services, and the quality and effectiveness of medical services are improved especially when these problems are solved successfully. A deep multitask and few-shot learning-based framework is proposed in this research. A mixed-expert neural network is adopted for the framework for multitask learning and predicting both likelihood of in-patient's mortality and length of stay. Additionally, the network architecture of the framework is used as an encoder for few-shot learning, enhancing the model's adaptability. Experimental results reveal a framework accuracy of 72.41%, slightly lower than the SVM model's 90% accuracy. However, the framework demonstrates a remarkable recall of 51.75%, which significantly exceeds the SVM model's recall of 4.35%. Moreover, it achieves an impressive F1-score of 60.40%, surpassing the performance of other models.