Natural Language Processing Basics

Abstract

Nephrology practice and research frequently require computational analysis of patient facts documented only in past clinical notes, ranging from the identification of basic facts such as whether the patient is thought to have CKD to prediction of patient outcomes such as kidney failure using phenotypes described in notes. As a result, researchers have long recognized the need for techniques to unlock facts described in notes for analysis and further facilitate clinical knowledge-sharing. Currently, natural language processing (NLP) techniques are facilitating care that would otherwise not be possible with standard structured data such as International Classification of Diseases (ICD) codes. This article reviews how NLP is used in nephrology and other fields of medicine to facilitate research, improve patient outcomes, and reduce physician workload. In addition, this article provides a high-level overview of the process to enable new collaborations between physicians and computer scientists. Motivations The first applications of clinical NLP stemmed out of specific information needs where it was clear that the answer lay within text already being captured and where provider workload could be reduced or the quality of care improved. For example, much work has gone into tasks requiring specific values such as populating the patient problem list,1 computer-assisted coding to identify appropriate billing codes,2 or search to identify patients matching specific criteria such as for a clinical trial or the identification of patients with CKD.3 More advanced applications leverage NLP as a component within more complex analyses such as classification or clustering analyses, or natural-language generation. Classifiers can be applied to various research tasks such as predicting survival of patients with CKD.4 Clustering algorithms can be particularly helpful at differentiating subclasses of a patient cohort, for example, the identification of subtypes of CKD.5 NLP is also used in bioinformatics to identify interactions between diseases, drugs, genes, and proteins documented in existing literature6 and to mine phenotypes from clinical notes used in genome-wide association studies (GWAS) and phenome-wide association studies analyses assessing gene function, as well as clinical applications predicting rare diseases.7 While many of these challenges are today best accomplished using an end-to-end machine learning (ML)–based approach, NLP has traditionally been a data acquisition step preceding statistical analysis. One early application was in MYCIN,8 an artificial intelligence (AI)‐based antibiotic recommender system. Where clinicians previously needed to answer a series of questions about the patient to get a recommendation, NLP was used to automatically obtain required patient information from notes. Approaches Although every NLP approach is different, they primarily fall into two categories: rule-based and ML-based approaches. Rule-based models are computationally simpler to execute and have been attempted since the early days of computer science. However, they require a lot of effort to perfect because it is up to the developers to program them to recognize every nuance of the language that affects meaning and map-related synonyms, abbreviations, and misspellings. ML-based NLP systems automatically learn how to solve a task from training data without being explicitly programmed to do so. This allows the computer to recognize statistically useful patterns of words that can later be used to predict a related result. With sufficient data, ML-based NLP can generally provide deeper insights and greater automation.9 As such, rule-based systems have been all but forgotten in fields with access to large volumes of training text. Comprehensive terminologies spanning all of medicine create powerful opportunities for enhancing NLP in medicine. Although physicians write notes using detailed descriptions, it is important for patients to be able to cluster around the higher-level implications of each feature. Consider a bottom-up study looking for consequences of a certain gene mutation that, unknown to the researchers, causes AKI symptoms. In order for the algorithm to properly cluster patients with AKI, it must understand that hemolytic uremic syndrome and milk alkali syndrome are both types of AKI, even if AKI is not explicitly mentioned in the notes. Conversely, an analysis of patients with AKI should also include patients with hemolytic uremic syndrome and milk alkali syndrome. It is possible that an ML system might recognize these relationships with enough data, but it is not guaranteed. Incorporating a knowledge base such as the Systematized Nomenclature of Medicine can resolve these challenges by allowing one to map from granular details to high-level concepts. Terminologies can also provide synonymous phrases that should be identified. To facilitate this form of analysis, most clinical NLP systems map to one or more terminologies. More recently, large repositories of unstructured data have been used to pretrain models that can readily be fine-tuned to particular tasks with minimal training data and improved generalization. For example, BioBERT,10 a domain-specific Bidirectional Encoder Representations from Transformers (BERT) model, is trained on large-scale biomedical corpora, and similarly, RadImageNet is pretrained with millions of radiologic images. Collaborating with Computer Scientists The application of ML to nephrology must begin with the nephrologist. However, given their medical duties and legal liabilities for the care being provided, physicians are understandably reluctant to take action on a patient on the basis of unexplained decisions made by black box algorithms. The success of any NLP-based study therefore requires strong interdisciplinary collaboration between medical experts and computer scientists. First, a goal needs to be formulated that is both clinically important and biologically plausible. For example, a nephrologist may note that being able to identify patients at risk of AKI would enable preemptive treatment and that AKI risk may in fact be estimated with reasonable accuracy on the basis of information available in electronic health records (EHRs). To improve the signal-to-noise ratio in the data and develop high-quality, reliable models, the medical expert can provide insight on what features of the data are most informative while avoiding confounders, errors, and unintentional biases. This is crucial for developing high-quality, reliable models. For example, rather than inundating an AKI detection model with all EHR data, it may help to select for ICD codes of conditions associated with higher risk of AKI along with relevant drugs such as aminoglycoside antibiotics, nonsteroidal anti-inflammatory drugs, and contrast agents. Finally, the expert should be closely involved during model development and validation to ensure quality and recognize errors and biases, lest the model fail in unexpected ways that would undermine trust in the health care system. For instance, in the Deepmind study,9 over 94% of the patients were male and AKI is known to have different patterns on the basis of sex. Race-based estimation of glomerular filtration rate was reported to overestimate kidney function in Black individuals and result in delayed treatment. The risk of unintentionally magnifying biases in data needs to be understood, and the ethics of applying ML should always be at the forefront when designing projects. Looking Ahead As long as medical notes remain the ground truth of patient facts captured at the point of care, NLP will play a critical role in deriving the best data to reduce practitioner workloads and improve clinical research. In an ever-growing and increasingly fragmented clinical environment, NLP can improve clinical knowledge management, facilitate new discoveries, and reduce physician workload and health care costs. Advances in ML, combined with national-level efforts to collect large medical datasets, will result in more generally applicable NLP tools facilitating novel applications; however, the application of NLP to medicine is unique because of the deep domain expertise needed and the importance of protecting patient privacy. With patient lives at stake, the cost of errors is incalculably high. For NLP and ML projects to be designed and implemented successfully, there must be deep interdisciplinary collaboration between physicians and computer scientists.

Topics & Concepts