DESIGNING PARAMETER AGNOSTIC FRAMEWORK FOR PREDICTING DISEASE ONSET USING DEEP LEARNING METHODS

Health Care and Life Sciences have been one of the most researched domains. The advent of newer computing paradigms has opened many new and unexplored avenues. These were first initiated by the newer instruments which further gave rise to newer paradigms that leverage the data generated using these instruments. With the HighPerformance Computing paradigms, becoming more accessible, we can now leverage various machine learning / deep learning approaches to utilize the data by digitizing it in various forms like Electronic Health Record (EHR) and Electronic Medical Record (EMR), etc. Mostly the medical notes comprise a lot of unstructured data and leveraging only structured information will lead to the loss of a significant amount of information. Although EHR contains a ton of data that can give valuable insights, it is still challenging to extract the information due to its inherent quality of being unstructured and missing data, variable length, and more importantly the lack of domain knowledge, which helps in the analysis. The Long Short-Term Memory (LSTMs) based implementation of Recurrent Neural Networks (RNNs) can be employed to solve the problem of variable-length records but the other issues like linking domain knowledge and extracting information from unstructured data remains a challenge. Also, many researchers have been employing several Deep Learning models to predict the onset of certain diseases by leveraging the data set pertaining to only the particular disease in question leaving the disease agnostic avenue for the trail. We endeavour to develop a generic framework that can be used to predict the onset of any disease by extracting valuable information out of unstructured medical data. In this work, we propose a generic framework that is not specific for any disease/disorder for prediction of the possibility of occurrence of a particular disease/disorder. We have employed RNNs with LSTMs to solve the variable-length issue and then leveraged the NLP based techniques to build a corpus to help map the domain-related information and also to extract the relevant information form the unstructured data. Our results show that the current model has a very small loss of 0.005 and average accuracy of 98.72 % with an AUC of 75.89%