Patient Similarity Model Using Discharge Sheet Representation and Final Diagnosis Prediction

Identifying similar patients is effective in designing many secondary applications to improve the quality of treatments and research services. The similarity of the final diagnoses is one of the aspects of similar patient groups. In order to measure similarity between patients, it is crucial to convert their information into a comparable format. There are different types of data in electronic health records (EHR). An important part of patient EHR are clinical notes, which face challenges to process. Therefore, the present study aims to design a clinical language processing model to identify definitive diagnoses.

In this study, the clinical notes of more than 26,000 patients from the MIMIC-III database were represented as vectors using modern language models, and these vectors were used as input for the diagnostic prediction model.

According to the results of the experiments, the BIO-BERT model with 0.715 and then the SciBERT model with 0.713 the best result between the biomedical language models. The results also show that using unique concepts extracted from clinical notes resulted in an increase in model accuracy.

Representation models trained with specific biomedical data can be used to map latent clinical note information to embedding vectors and provide the ability to use notes in machine learning algorithms, including prediction of the final diagnostic group.