Rasool Samani (رسول سامانی)

MSc student in software engineering and data science

پيش بيني پذيرش مجدد بيمار با استفاده از بازنمايي پرونده الكترونيك سلامت و گراف ناهمگون مفاهيم پزشكي

Predicting Patient Readmission Using Electronic Health Record Representation and Heterogeneous Graph of Medical Concepts

The representation of electronic health records is a crucial and fundamental topic in analyzing medical and biomedical data in computer science. The electronic health record contains diverse and significant data from the patient’s treatment process, including structured data such as medical codes,  semistructured data such as laboratory results, and unstructured data such as clinical reports. We exploited the extraction of appropriate features and modeling of this data to improve the accuracy of various tasks related to medical text mining, such as predicting patient readmission, predicting patient death, and estimating the duration of hospitalization.

In recent years, several research works are conducted to propose representations of electronic health records, each addressing a different aspect. These studies, however, used only the texts of the patient’s clinical reports or only the medical codes from the patient’s EHR to create representations which led to the loss of critical information from other parts of the electronic health record.

In this study, we proposed a model by exploiting the combination of electronic health record representation and embedding patients ’communication graphs. First of all, we covered as much as possible electronic health record data in our representation, and then to find similar pairs, we built patients’ relational graphs. In this regard, a heterogeneous graph of patients’ information, including medical codes and drugs, was generated. Appropriate feature vectors were created by graph embedding techniques and combined with the vectors prepared from the patient’s electronic record representation. Comparing the efficiency and accuracy of this approach with other related state-of-the-art works shows that the proposed model outperforms other patient readmission classification tasks and achieves an accuracy of 72.2 with the AUROC criterion.