Detection of Bleeding Events In EHR Notes Using A Graph Dynamic Autoencoder Based Riemannian Residual Neural Network

Identifying occurrences of abnormal bleeding based on unstructured and structured medical data, such as patient notes, vitals, and laboratory findings, is the problem of recognising bleeding events in Electronic Health Records (EHR). The problem is in accurately recognising these occurrences by analysing temporal patterns, extracting significant information, and distinguishing catastrophic bleeding occasions from typical medical fluctuations given the complexity of EHR data, which is frequently longitudinal, heterogeneous, and packed with noise. The goal is to create a Graph Dynamic Autoencoder combined with a Riemannian Residual Neural Network (GDA-RRNN) with Elk Herd optimizer (EHO) model that can identify and display bleeding incidents in electronic medical data. The EHO optimizes the hyperparameters GDA-RRNN to improve the performance of the suggested method. We extracted 300 electronic health records that contained leukaemia or bleeding diagnostic codes from the International categorization of Diseases, Tenth Revision. The electronic health record has annotations indicating whether a sentence was positive or negative for bleeding. A proposed model that can identify bleeding at the sentence and note levels was created using the annotated texts. The best-performing suggested approach achieved a sensitivity of 0.99, specificity of 0.99, and a negative prediction value of 0.99 on an equal test set of 1178 phrases.