Disease Inference from Health-Related Questions using Sparse Neural Network and LSTM

Abstract

There are many relationship and context retrieval related problems that are resolved using medical records with Deep learning-based techniques, but when it comes to community data such as health care forums where the quality of data cannot be met as health records due to gap in medical vocabulary it becomes impossible to provide an accurate solution, our research involves in providing a solution to this problem. The graph algorithms have always provided the best solutions to map and normalize in NLP domain, we have used the same to find the normalized medical terms for out user questions in the health care forums. Instead of training the actual data from the forums with the LSTM, we create medical signatures of the words coming together to form a context in the medical dictionary. We consider the words used in the dataset as vertices and find dense subgraphs to uniquely identify the condition with medical dictionary data. In simple words, we aim to build a system to convert the vague descriptions of the disease to match the accurate medical term from the medical dictionary such as snomedCT. We used the words which co-occur to define our relations which will, in turn, provide us with a solution to bridge the gap of medical vocabulary. The mappings of normalized terms are foundations to build the hidden layer of our neural networks, instead of constructing a direct connection between the input neurons to all hidden neurons we connect only the subgraph results thus improving our accuracy to a better level then existing methodologies.