Learning distributional representation of clinical concepts (e.g., diseases, drugs, and labs) is an important research area of deep learning in the medical domain. However, many existing relevant methods do not consider temporal dependencies along the longitudinal sequence of a patient’s records, which may lead to incorrect selection of contexts. To address this issue, we extended three popular concept embedding learning methods: word2vec, positive pointwise mutual information (PPMI) and FastText, to consider time-sensitive information. We then trained them on a large electronic health records (EHR) database containing about 50 million patients to generate concept embeddings and evaluated them for both intrinsic evaluations focusing on concept similarity measure and an extrinsic evaluation to assess the use of generated concept embeddings in the task of predicting disease onset. Our experiments show that embeddings learned from information within one visit (time window zero) improve performance on the concept similarity measure and the FastText algorithm usually had better performance than the other two algorithms. For the predictive modeling task, the optimal result was achieved by word2vec embeddings with a 30-day sliding window. Considering time constraints are important in training clinical concept embeddings. We expect they can benefit a series of downstream applications.

The COVID-19 pandemic resulted in a large expansion of telehealth, but little is known about user sentiment. Tweets containing the terms “telehealth” and “telemedicine” were extracted (n = 192,430) from the official Twitter API between November 2019 and April 2020. A random subset of 2000 tweets was annotated by trained readers to classify tweets according to their content, including telehealth, sentiment, user type, and relation to COVID-19. A state-of-the-art NLP model (Bidirectional Encoder Representations from Transformers, BERT) was used to categorize the remaining tweets. Following a low and fairly stable level of activity, telehealth tweets rose dramatically beginning the first week of March 2020. The sentiment was overwhelmingly positive or neutral, with only a small percentage of negative tweets. Users included patients, clinicians, vendors (entities that promote the use of telehealth technology or services), and others, which represented the largest category. No significant differences were seen in sentiment across user groups. The COVID-19 pandemic produced a large increase in user tweets related to telehealth and COVID-19, and user sentiment suggests that most people feel positive or neutral about telehealth

We present a Three-level Hierarchical Transformer Network (3-level-HTN) for modeling long-term dependencies across clinical notes for the purpose of patient-level prediction. The network is equipped with three levels of Transformer-based encoders to learn progressively from words to sentences, sentences to notes, and finally notes to patients. The first level from word to sentence directly applies a pre-trained BERT model as a fully trainable component. While the second and third levels both implement a stack of transformer-based encoders, before the final patient representation is fed into a classification layer for clinical predictions. Compared to conventional BERT models, our model increases the maximum input length from 512 tokens to much longer sequences that are appropriate for modeling large numbers of clinical notes. We empirically examine different hyper-parameters to identify an optimal trade-off given computational resource limits. Our experiment results on the MIMIC-III dataset for different prediction tasks demonstrate that the proposed Hierarchical Transformer Network outperforms previous state-of-the-art models, including but not limited to BigBird.

Neural network-based representations ("embeddings") have dramatically advanced natural language processing (NLP) tasks, including clinical NLP tasks such as concept extraction. Recently, however, more advanced embedding methods and representations (e.g., ELMo, BERT) have further pushed the state-of-the-art in NLP, yet there are no common best practices for how to integrate these representations into clinical tasks. The purpose of this study, then, is to explore the space of possible options in utilizing these new models for clinical concept extraction, including comparing these to traditional word embedding methods (word2vec, GloVe, fastText). Both off-the-shelf open-domain embeddings and pre-trained clinical embeddings from MIMIC-III are evaluated. We explore a battery of embedding methods consisting of traditional word embeddings and contextual embeddings, and compare these on four concept extraction corpora: i2b2 2010, i2b2 2012, SemEval 2014, and SemEval 2015. We also analyze the impact of the pre-training time of a large language model like ELMo or BERT on the extraction performance. Last, we present an intuitive way to understand the semantic information encoded by contextual embeddings. Contextual embeddings pre-trained on a large clinical corpus achieves new state-of-the-art performances across all concept extraction tasks. The best-performing model outperforms all state-of-the-art methods with respective F1-measures of 90.25, 93.18 (partial), 80.74, and 81.65. We demonstrate the potential of contextual embeddings through the state-of-the-art performance these methods achieve on clinical concept extraction. Additionally, we demonstrate contextual embeddings encode valuable semantic information not accounted for in traditional word representations.