Latent topic discovery of clinical concepts from hospital discharge summaries of a heterogeneous patient cohort

Making sense of abbreviations in nursing notes: A case study on mortality prediction

Unstructured data from electronic health records hold potential for improving predictive models for health outcomes. Efforts to extract structured information from the unstructured data used text mining methodologies, such as topic modeling and sentiment analysis. However, such methods do not account for abbreviations. Nursing notes have valuable information about nurses' assessments and interventions, and the abbreviation use is common. Thus, abbreviation disambiguation may add more insight when using unstructured text for predictive modeling. We present a new process to extract structured information from nursing notes through abbreviation normalization, lemmatization, and stop word removal. Our study found that abbreviation disambiguation in nursing notes for subsequent topic modeling and sentiment analysis improved prediction of in-hospital and 30-day mortality while controlling for comorbidity.

Discovering associations between problem list and practice setting

Background

The Health Information Technology for Economic and Clinical Health Act (HITECH) has greatly accelerated the adoption of electronic health records (EHRs) with the promise of better clinical decisions and patients’ outcomes. One of the core criteria for “Meaningful Use” of EHRs is to have a problem list that shows the most important health problems faced by a patient. The implementation of problem lists in EHRs has a potential to help practitioners to provide customized care to patients. However, it remains an open question on how to leverage problem lists in different practice settings to provide tailored care, of which the bottleneck lies in the associations between problem list and practice setting.

Methods

In this study, using sampled clinical documents associated with a cohort of patients who received their primary care at Mayo Clinic, we investigated the associations between problem list and practice setting through natural language processing (NLP) and topic modeling techniques. Specifically, after practice settings and problem lists were normalized, statistical χ² test, term frequency-inverse document frequency (TF-IDF) and enrichment analysis were used to choose representative concepts for each setting. Then Latent Dirichlet Allocations (LDA) were used to train topic models and predict potential practice settings using similarity metrics based on the problem concepts representative of practice settings. Evaluation was conducted through 5-fold cross validation and Recall@k, Precision@k and F1@k were calculated.

Results

Our method can generate prioritized and meaningful problem lists corresponding to specific practice settings. For practice setting prediction, recall increases from 0.719 (k = 2) to 0.931 (k = 10), precision increases from 0.882 (k = 2) to 0.931 (k = 10) and F1 increases from 0.790 (k = 2) to 0.931 (k = 10).

Conclusion

To our best knowledge, our study is the first attempting to discover the association between the problem lists and hospital practice settings. In the future, we plan to investigate how to provide more tailored care by utilizing the association between problem list and practice setting revealed in this study.

Electronic supplementary material

The online version of this article (10.1186/s12911-019-0779-y) contains supplementary material, which is available to authorized users.

Estimating Patient's Health State Using Latent Structure Inferred from Clinical Time Series and Text

Modern intensive care units (ICUs) collect large volumes of data in monitoring critically ill patients. Clinicians in the ICUs face the challenge of interpreting large volumes of high-dimensional data to diagnose and treat patients. In this work, we explore the use of Hierarchical Dirichlet Processes (HDP) as a Bayesian nonparametric framework to infer patients' states of health by combining multiple sources of data. In particular, we employ HDP to combine clinical time series and text from the nursing progress notes in a probabilistic topic modeling framework for patient risk stratification. Given a patient cohort, we use HDP to infer latent "topics" shared across multimodal patient data from the entire cohort. Each topic is modeled as a multinomial distribution over a vocabulary of codewords, defined over heterogeneous data sources. We evaluate the clinical utility of the learned topic structure using the first 24-hour ICU data from over 17,000 adult patients in the MIMIC-II database to estimate patients' risks of in-hospital mortality. Our results demonstrate that our approach provides a viable framework for combining different data modalities to model patient's states of health, and can potentially be used to generate alerts to identify patients at high risk of hospital mortality.

Phenotyping Hypotensive Patients in Critical Care Using Hospital Discharge Summaries

Among critically-ill patients, hypotension represents a failure in compensatory mechanisms and may lead to organ hypoperfusion and failure. In this work, we adopt a data-driven approach for phenotype discovery and visualization of patient similarity and cohort structure in the intensive care unit (ICU). We used Hierarchical Dirichlet Process (HDP) as a nonparametric topic modeling technique to automatically learn a d-dimensional feature representation of patients that captures the latent "topic" structure of diseases, symptoms, medications, and findings documented in hospital discharge summaries. We then used the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm to convert the d-dimensional latent structure learned from HDP into a matrix of pairwise similarities for visualizing patient similarity and cohort structure. Using discharge summaries of a large patient cohort from the MIMIC II database, we evaluated the clinical utility of the discovered topic structure in phenotyping critically-ill patients who experienced hypotensive episodes. Our results indicate that the approach is able to reveal clinically interpretable clustering structure within our cohort and may potentially provide valuable insights to better understand the association between disease phenotypes and outcomes.