مدل سازی شباهت بیمار با استفاده از بازنمایی هوشمند خلاصه پرونده برای پیش بینی تشخیص نهایی.

Introduction: The clinical trials recorded in the electronic health record (EHR) contains important information about the patient's history and the treatments performed. Since clinical notes are stored unstructured, they cannot be applied directly in machine learning algorithms. One way to structure textual data is to represent them as vectors. Methods: In this research, the discharge sheets are used to generate the vector corresponding for each patient. Language models are used to represent the latest text processing methods. The dataset contains the discharge sheets of more than 26,000 patient records from the Medical Information Mart for Intensive Care III (MIMIC-III) database. To analyze the quality of representation framework, the diagnosis prediction downstream task is used and the evaluation criteria are reported for each language model. Results: Among the LLMs used in the framework, the best one for the discharge sheets is the BIO-BERT model and then the SciBERT model, which produced the ROC_AUC 0.715 and 0.713 respectively. This evaluation criterion is used to check the quality of forecasting models. The use of clinical text preprocessing and mapping of clinical entities to their standard names in the UMLS knowledge base has improved the evaluation criteria for specific language models in the biomedical field, and the greatest improvement is related to the UMLSBERT model, which is trained on the standard names of the knowledge base. Conclusion: BIO-BERT and SciBERT language models that trained on the data of clinical papers are suggested as the best option for representing the discharge sheet to vectors. However, since the for discharge sheets are different from the scientific paper in terms of structure and content, the preprocessing of clinical trials in order to identify entities and map them to knowledge sources to fetch the standard names of clinical concepts improves the results obtained in clinical LLMs. [ABSTRACT FROM AUTHOR]