With the abundance of natural language processing (NLP) frameworks and toolkits being used in the clinical arena, a new challenge has arisen - how do technologists collaborate across several projects in an easy way? Private sector companies are usually not willing to share their work due to intellectual property rights and profit-bearing decisions. Therefore, the annotation schemes and toolkits that they use are rarely shared with the wider community. We present the clinical language pipeline toolkit (CLPT) and its corresponding annotation scheme called the CLAO (Clinical Language Annotation Object) with the aim of creating a way to share research results and other efforts through a software solution. The CLAO is a unified annotation scheme for clinical technology processing (CTP) projects that forms part of the CLPT and is more reliable than previous standards such as UIMA, BioC, and cTakes for annotation searches, insertions, and deletions. Additionally, it offers a standardized object that can be exchanged through an API that the authors release publicly for CTP project inclusion.

Automatically classifying electronic health records (EHRs) into diagnostic codes has been challenging to the NLP community. State-of-the-art methods treated this problem as a multi-label classification problem and proposed various architectures to model this problem. However, these systems did not leverage the superb performance of pretrained language models, which achieved superb performance on natural language understanding tasks. Prior work has shown that pretrained language models underperformed on this task with the regular fine-tuning scheme. Therefore, this paper aims at analyzing the causes of the underperformance and developing a framework for automatic ICD coding with pretrained language models. We spotted three main issues through the experiments: 1) large label space, 2) long input sequences, and 3) domain mismatch between pretraining and fine-tuning. We propose PLM-ICD, a framework that tackles the challenges with various strategies. The experimental results show that our proposed framework can overcome the challenges and achieves state-of-the-art performance in terms of multiple metrics on the benchmark MIMIC data. Our source code is available at https://github.com/MiuLab/PLM-ICD.

Acronym disambiguation (AD) is the process of identifying the correct expansion of the acronyms in text. AD is crucial in natural language understanding of scientific and medical documents due to the high prevalence of technical acronyms and the possible expansions. Given that natural language is often ambiguous with more than one meaning for words, identifying the correct expansion for acronyms requires learning of effective representations for words, phrases, acronyms, and abbreviations based on their context. In this paper, we proposed an approach to leverage the triplet networks and triplet loss which learns better representations of text through distance comparisons of embeddings. We tested both the triplet network-based method and the modified triplet network-based method with m networks on the AD dataset from the SDU@AAAI-21 AD task, CASI dataset, and MeDAL dataset. F scores of 87.31%, 70.67%, and 75.75% were achieved by the m network-based approach for SDU, CASI, and MeDAL datasets respectively indicating that triplet network-based methods have comparable performance but with only 12% of the number of parameters in the baseline method. This effective implementation is available at https://github.com/sandaruSen/m_networks under the MIT license.

Writing the conclusion section of radiology reports is essential for communicating the radiology findings and its assessment to physician in a condensed form. In this work, we employ a transformer-based Seq2Seq model for generating the conclusion section of German radiology reports. The model is initialized with the pretrained parameters of a German BERT model and fine-tuned in our downstream task on our domain data. We proposed two strategies to improve the factual correctness of the model. In the first method, next to the abstractive learning objective, we introduce an extraction learning objective to train the decoder in the model to both generate one summary sequence and extract the key findings from the source input. The second approach is to integrate the pointer mechanism into the transformer-based Seq2Seq model. The pointer network helps the Seq2Seq model to choose between generating tokens from the vocabulary or copying parts from the source input during generation. The results of the automatic and human evaluations show that the enhanced Seq2Seq model is capable of generating human-like radiology conclusions and that the improved models effectively reduce the factual errors in the generations despite the small amount of training data.

Radiology report is an official record of radiologists’ interpretation of patients’ radiographs and it’s a crucial component in the overall medical diagnostic process. However, it can contain various types of errors that can lead to inadequate treatment or delay in diagnosis. To address this problem, we propose a deep learning framework to detect errors in radiology reports. Specifically, our method detects errors between findings and conclusion of chest X-ray reports based on a supervised learning framework. To compensate for the lack of data availability of radiology reports with errors, we develop an error generator to systematically create artificial errors in existing reports. In addition, we introduce a Medical Knowledge-enhancing Pre-training to further utilize the knowledge of abbreviations and key phrases frequently used in the medical domain. We believe that this is the first work to propose a deep learning framework for detecting errors in radiology reports based on a rich contextual and medical understanding. Validation on our radiologist-synthesized dataset, based on MIMIC-CXR, shows 0.80 and 0.95 of the area under precision-recall curve (AUPRC) and the area under the ROC curve (AUROC) respectively, indicating that our framework can effectively detect errors in the real-world radiology reports.

In this work, cross-linguistic span prediction based on contextualized word embedding models is used together with neural machine translation (NMT) to transfer and apply the state-of-the-art models in natural language processing (NLP) to a low-resource language clinical corpus. Two directions are evaluated: (a) English models can be applied to translated texts to subsequently transfer the predicted annotations to the source language and (b) existing high-quality annotations can be transferred beyond translation and then used to train NLP models in the target language. Effectiveness and loss of transmission is evaluated using the German Berlin-Tübingen-Oncology Corpus (BRONCO) dataset with transferred external data from NCBI disease, SemEval-2013 drug-drug interaction (DDI) and i2b2/VA 2010 data. The use of English models for translated clinical texts has always involved attempts to take full advantage of the benefits associated with them (large pre-trained biomedical word embeddings). To improve advances in this area, we provide a general-purpose pipeline to transfer any annotated BRAT or CoNLL format to various target languages. For the entity class medication, good results were obtained with 0.806 F1-score after re-alignment. Limited success occurred in the diagnosis and treatment class with results just below 0.5 F1-score due to differences in annotation guidelines.

Decision support systems based on clinical notes have the potential to improve patient care by pointing doctors towards overseen risks. Predicting a patient’s outcome is an essential part of such systems, for which the use of deep neural networks has shown promising results. However, the patterns learned by these networks are mostly opaque and previous work revealed both reproduction of systemic biases and unexpected behavior for out-of-distribution patients. For application in clinical practice it is crucial to be aware of such behavior. We thus introduce a testing framework that evaluates clinical models regarding certain changes in the input. The framework helps to understand learned patterns and their influence on model decisions. In this work, we apply it to analyse the change in behavior with regard to the patient characteristics gender, age and ethnicity. Our evaluation of three current clinical NLP models demonstrates the concrete effects of these characteristics on the models’ decisions. They show that model behavior varies drastically even when fine-tuned on the same data with similar AUROC score. These results exemplify the need for a broader communication of model behavior in the clinical domain.

Existing question answering (QA) datasets derived from electronic health records (EHR) are artificially generated and consequently fail to capture realistic physician information needs. We present Discharge Summary Clinical Questions (DiSCQ), a newly curated question dataset composed of 2,000+ questions paired with the snippets of text (triggers) that prompted each question. The questions are generated by medical experts from 100+ MIMIC-III discharge summaries. We analyze this dataset to characterize the types of information sought by medical experts. We also train baseline models for trigger detection and question generation (QG), paired with unsupervised answer retrieval over EHRs. Our baseline model is able to generate high quality questions in over 62% of cases when prompted with human selected triggers. We release this dataset (and all code to reproduce baseline model results) to facilitate further research into realistic clinical QA and QG: https://github.com/elehman16/discq.

Word embeddings have been widely used in Natural Language Processing (NLP) tasks. Although these representations can capture the semantic information of words, they cannot learn the sequence-level semantics. This problem can be handled using contextual word embeddings derived from pre-trained language models, which have contributed to significant improvements in several NLP tasks. Further improvements are achieved when pre-training these models on domain-specific corpora. In this paper, we introduce Clinical Flair, a domain-specific language model trained on Spanish clinical narratives. To validate the quality of the contextual representations retrieved from our model, we tested them on four named entity recognition datasets belonging to the clinical and biomedical domains. Our experiments confirm that incorporating domain-specific embeddings into classical sequence labeling architectures improves model performance dramatically compared to general-domain embeddings, demonstrating the importance of having these resources available.

Recent studies show that neural natural processing models for medical code prediction suffer from a label imbalance issue. This study aims to investigate further imbalance in a medical code prediction dataset in terms of demographic variables and analyse performance differences in demographic groups. We use sample-based metrics to correctly evaluate the performance in terms of the data subject. Also, a simple label distance metric is proposed to quantify the difference in the label distribution between a group and the entire data. Our analysis results reveal that the model performs differently towards different demographic groups: significant differences between age groups and between insurance types are observed. Interestingly, we found a weak positive correlation between the number of training data of the group and the performance of the group. However, a strong negative correlation between the label distance of the group and the performance of the group is observed. This result suggests that the model tends to perform poorly in the group whose label distribution is different from the global label distribution of the training data set. Further analysis of the model performance is required to identify the cause of these differences and to improve the model building.

In this paper, we investigate ensemble methods for fine-tuning transformer-based pretrained models for clinical natural language processing tasks, specifically temporal relation extraction from the clinical narrative. Our experimental results on the THYME data show that ensembling as a fine-tuning strategy can further boost model performance over single learners optimized for hyperparameters. Dynamic snapshot ensembling is particularly beneficial as it fine-tunes a wide array of parameters and results in a 2.8% absolute improvement in F1 over the base single learner.

Sequence-to-sequence models are appealing because they allow both encoder and decoder to be shared across many tasks by formulating those tasks as text-to-text problems. Despite recently reported successes of such models, we find that engineering input/output representations for such text-to-text models is challenging. On the Clinical TempEval 2016 relation extraction task, the most natural choice of output representations, where relations are spelled out in simple predicate logic statements, did not lead to good performance. We explore a variety of input/output representations, with the most successful prompting one event at a time, and achieving results competitive with standard pairwise temporal relation extraction systems.