Classifying Implant-Bearing Patients via their Medical Histories: a Pre-Study on Swedish EMRs with Semi-Supervised GanBERT

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5428–5435 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5428 Classifying Implant-Bearing Patients via their Medical Histories: a Pre-Study on Swedish EMRs with Semi-Supervised GAN-BERT Benjamin Danielsson∗, Marina Santini§, Peter Lundberg†, Yosef Al-Abasse∗∗, Arne J¨onsson∗, Emma Eneling‡, Magnus Stridsman‡ ∗Department of Computer and Information Science, Link¨oping University, Sweden benda425@student.liu.se, arne.jonsson@liu.se §RISE Research Institutes of Sweden, Sweden marina.santini@ri.se †Radiation Physics, Center for Medical Imaging and Visualization (CMIV), Link¨oping University, and Department of Health, Medicine and Care, Link¨oping University, Link¨oping, Sweden peter.lundberg@liu.se ∗∗Radiation Physics, and Department of Health, Medicine and Care, Link¨oping University, Link¨oping, Sweden yosef.al-abasse@regionostergotland.se ‡Unit for Technology Assessment, Testing and Innovation, and Department of Biomedical Engineering, Link¨oping University, Link¨oping, Sweden emma.eneling@regionostergotland.se, magnus.stridsman@regionostergotland.se Abstract In this paper, we compare the performance of two BERT-based text classifiers whose task is to classify patients (more precisely, their medical histories) as having or not having implant(s) in their body. One classifier is a fully-supervised BERT classifier. The other one is a semi-supervised GAN-BERT classifier. Both models are compared against a fully-supervised SVM classifier. Since fully-supervised classification is expensive in terms of data annotation, with the experiments presented in this paper, we investigate whether we can achieve a competitive performance with a semi-supervised classifier based only on a small amount of annotated data. Results are promising and show that the semi-supervised classifier has a competitive performance when compared with the fully-supervised classifier. Keywords: text classification, BERT, GAN-BERT, electronic medical records, EMR, clinical text mining 1. Introduction The paradigm shift in NLP (Natural Language Pro- cessing) (Sun et al., 2021) based on large pre-trained language models that can be fine-tuned to downstream tasks has boosted the performance of many NLP tasks when tested on traditional benchmarks (Min et al., 2021). In the experiments presented in this paper, we investi- gate a downstream task that is related to a real-world need, where standard benchmarks are not available, namely the automatic identification of patients who bear implant(s) in their body. Identifying the presence of implants in certain patients is important for radiologists and other clinical profes- sionals because some implants are not compatible with MRI scanning. The current process to ascertain the presence of implants in a patient is manual (Kihlberg and Lundberg, 2019), slow and error prone, especially if patients are elderly and might have forgotten about implants they had when they were younger. Addition- ally, even if implants are removed, parts (like leads) might remain in the body and cause damage (burns or scorches) to the patient during the MRI scanning. The core idea of these experiments is then to classify implant-bearing patients on the basis of their medical histories. We rely on the following assumption: if pa- tients underwent an implant operation, certainly this in- formation is stored in patients’ medical records. By sieving through medical records, we circumvent the risk of overlooking crucial clinical information that are useful for professional staff. Therefore, rather than relying on …

Content truncated to 500 words