Deep Learning models for patient similarity analysis in personalized medicine

Abstract

Mining Electronic Health Records (EHR) is more complex than standard data mining operations due to their noisy, irregular, and diversified nature. An overall model is a useful tool for disease prediction that is, using all available training data to construct a global model and then using this model to forecast health risk for each patient. The benefit of adopting a single model is that it captures all the information from the training population as a whole. Patients may have a variety of phenotype, medical conditions, and so forth. When using a global model, important unique information that is crucial for particular patients may be excluded. As a result, personalized therapy requires the development of a targeted, patient-specific model for each individual patient. newlineThe proposed approach builds a single class of flexible, deep representations that may be used in a variety of ways, including predicting different outcomes and using them as patient similarity measurements. These advantageous applications in the healthcare environment are not only utilized, but also serve as a way of more thoroughly examining the usability of learned representations. Initially, a Convolutional Neural Network (CNN) is used along with Triplet-Mining Metric Learning (TMML) to analyse EHRs that include crucial local information and when the training is complete, the distance and similarity scores are computed. This similarity information is used to make disease forecasts and categorize patients effectively. newline newline