Application of Natural Language Processing in Nephrology Research.

Electronic health records (EHRs) are prevalent throughout the United States. EHR data are stored in structured (e.g., vital signs and laboratory results) and unstructured (e.g., progress notes and radiology reports) formats. Unstructured data contain critical information regarding patients' health and behavior. However, unstructured data cannot be used in its raw form and traditionally require manual chart abstraction, which is time and labor intensive. Natural language processing (NLP) is a branch of artificial intelligence that enables computers to understand human language. NLP allows researchers to collect data from the free text of EHRs and other free text, providing another source of data for researchers. This will allow for the identification of patients with certain diseases or disease phenotypes, which can enhance patient recruitment into clinical trials. In this review, we will provide examples of how NLP has been used in nephrology research. We focus our review on identification of CKD, hypertension, symptoms, inclusion of NLP features to improve risk prediction, and identifying trending topics in research. Identification of Patients with Specific Phenotypes CKD is under-recognized by primary care providers, and referral to nephrology care is often delayed. Early referral to nephrology allows for risk-factor mitigation in addition to implementation of novel pharmacotherapies shown to decrease the progression of kidney disease. Chase et al. used two methods of NLP, a lexical-based classifier and a word-count method, to assess whether CKD was properly documented in the medical records.1 First, authors developed a library of terms associated with CKD using notes written by kidney specialists. Then, association testing was performed to identify terms that had high attributable strength with CKD. The lexical-based classifier was trained using notes of patients with known CKD status and tested in a separate cohort of patients. The word-count method simply looks for the presence or absence of a CKD term and categorizes it as positive if there is even one instance. Both the lexical-based classifier and word-count method were extremely sensitive (95.8% and 99.8%, respectively) and specific (99.8% and 98.8%, respectively) for determining CKD documentation. In addition, it was found that 22% of patients with moderate CKD did not have it documented. The authors observed that patients without CKD documented were less likely to be on an ACE inhibitor/angiotensin receptor blocker and were less likely to have their urine protein evaluated, potentially indicating that a lack of documentation may represent a lack of recognition. This showcases NLP's potential to alert physicians to a diagnosis of CKD and thus lead to appropriate management. Hypertension management is crucial for prevention of CKD progression and cardiovascular complications. BP readings may be located in multiple locations in the EHR, including the formal flowsheet but also located as free text within notes. Greenberg et al. explored using regular expressions as a way to increase the yield of detection for BP control of their nephrology clinic panel.2 Regular expressions are sequences of letters or patterns that are used to identify a concept, in this case BP readings. The authors found that only 37% of their patients were considered under control using the most recent flowsheet BP, and using more than one BP reading increased their proportion controlled to 42.3%. The authors were able to increase their overall yield of BP readings after using NLP to extract free-text BP readings and found that 52.6% of their panel was controlled. This study highlights NLP's utility in detecting BP data that may go unnoticed and its ability to improve clinical quality improvement assessment. Patient-Centered Outcomes Research Patient symptoms are outcomes identified as important to patients and caregivers.3 Unfortunately, patient symptoms are poorly captured in the structured portion of the EHR but frequently documented in progress notes as free text. In our prior work, we used NLP to identify seven key symptoms from two EHRs of patients on hemodialysis (HD).4 Common symptoms were pain, fatigue, and nausea and/or vomiting. The NLP algorithm here parsed words from progress notes to Systematized Nomenclature of Medicine Clinical Terms clinical terms, a library of clinical terms.5 NLP outperformed International Classification of Diseases codes in regards to sensitivity and negative predictive value and had similar specificity and positive predictive value for all seven symptoms. The results were similar across an internal validation and external validation cohort. One of the first steps to improving the undertreatment of patient symptoms is to increase recognition of said symptoms. This method is one possible way of improving recognition. One limitation of this study was the use of manual chart review as the reference standard. A follow-up study using patient surveys as the reference standard is currently underway. In addition to extraction of specific symptoms, NLP can also be used for sentiment analysis. Sentiment analysis uses NLP, machine learning, and statistics to determine the emotions contained within the text. Huang et al. developed an application where patients on home HD could report emotional status using a sliding scale and a free-text field where patients could write a note about the session.6 They used a previously developed classifier to classify the patient's free text as positive, negative, or neutral. Of the 9379 sessions with session notes, 19% were classified as positive, 16% were negative, and the remaining were neutral. On deeper analysis, researchers found that personal health and emotions were frequently observed. Because home dialysis patients are only seen monthly, this type of remote monitoring allows for identifying patients who are struggling with home dialysis. Patients with repeated negative sentiments can be flagged for physician review to address potentially modifiable treatment parameters and reduce transitions to in-center HD. Improving Risk Prediction The Tangri risk score is a well-validated score using readily available structured clinical data to predict a patient's 5-year risk for developing kidney failure.7 NLP unlocks data within the EHR that may allow us to further improve clinical predictors. Perotte et al. tested various prediction models for progression of CKD3 to CKD4.8 In two of the five models, they included features extracted from clinical documents using Latent Dirichlet Allocation (LDA). LDA is a topic modeling approach where each document is represented by some number of topics, and each topic is represented by a set of words. Of the models tested, the time-series model, which incorporated both laboratory data and concepts from clinical note, was statistically significantly better than the other models that included only laboratory data (concordance 0.849). The ability to accurately predict progression to CKD4 may allow for improved early identification of high-risk patients and prompt referral to a nephrologist. AKI is common and associated with poor clinical outcomes. There is limited ability to prevent AKI, but there is greater potential to create and test therapeutics with better predictors. Sun et al. used typical structured data including vital signs and laboratory markers, in conjunction with clinical notes from the Medical Information Mart for Intensive Care-III database, to predict AKI onset 72 hours from ICU admission.9 The authors used a bag-of-words approach to create data from the notes. They identified which words were present in the text, removed low-frequency words, and removed stop words as per the National Center for Biotechnology Information guide. Researchers tested four different prediction models comparing structured data, unstructured data, and both. Ultimately, prediction models using both structured and unstructured data produced the best area under the receiver operating characteristic curve (0.8352), while the best area under the receiver operating characteristic curve produced by structured data alone was 0.8132. Features with high impact on models included words such as “Lasix” and “cognitive impairment.” This study provides another example of how inclusion of features available in free text can improve performance of models. Despite a shortage of kidneys available for kidney transplantation, not all kidneys are suitable for transplantation. Placona et al. used NLP to build a model to predict the probability of delay or discard for adult deceased kidney donors using unstructured data.10 Their NLP model was similar to models built using structured data alone. Whether addition of NLP features improved discrimination of a model built using structured data alone was not tested. NLP for Nephrology Education Identifying topics of interest and changes in trending topics in nephrology research over time can be challenging. NLP can aid in digesting the large amount of nephrology research topics. Zengul et al. used various text mining techniques, including NLP, to perform topic analysis from the abstracts of the top ten journals in nephrology.11 After indexing the abstracts that met their inclusion and exclusion criteria, researchers had to generate term lists and curate the phrases for the NLP algorithm to parse. This process included an automated portion that used a curated medical terminology list and a manual portion of adding phrases believed to be relevant and important. They identified ten distinct categories from 17,000 abstracts. They were also able to identify trends in topics and noted the increasing trend in publication on patient-related outcomes and perspectives of clinicians. This analysis can be informative for educators and clinicians to aid in keeping informed of trending topics. Limitations While NLP has many potential uses, there are several limitations. Physician bias in the form of stigmatizing language in the EHR and selective screening for adverse behaviors can get perpetuated by NLP. Current studies lack external validation and may not generalize given regional and hospital system differences in language and documentation. In this review, we provided examples of how NLP has been used in nephrology research. NLP can identify patients with specific disease phenotypes through unstructured data in the EHR, with abilities to contribute to practice quality improvement, symptom outcomes, risk prediction, and medical education. Given these promising studies, hopefully trials evaluating the clinical implementation of NLP into clinical care are forthcoming.