Abstract
Background
Syncope is a commonly occurring presenting symptom in emergency departments. While the majority of episodes are benign, syncope is associated with worse prognosis in hypertrophic cardiomyopathy, arrhythmia syndromes, heart failure, aortic stenosis and coronary heart disease. Flagging documented syncope in these patients may be crucial to management decisions.
Previous studies show that the International Classification of Diseases (ICD) codes for syncope have a sensitivity of around 0.63, leading to a large number of false negatives if patient identification is based on administrative codes. Thus, in order to provide data-driven, clinical decision support, and to improve identification of patient cohorts for research, better tools are needed.
A recent study manually annotated more than 30.000 patient records in order to develop a natural language processing (NLP) tool, which achieved a sensitivity of 92.2%. Since access to medical records and annotation resources is limited, we aimed to investigate whether an unsupervised machine learning and NLP approach with no manual input could achieve similar performance.
Methods
Our data was admission notes for adult patients admitted between 2005 and 2016 at a large university hospital in Norway. 500 records from patients with, and 500 without a “R55 Syncope” ICD code at discharge were drawn at random. R55 code was considered “ground truth”. Headers containing information about tentative diagnoses were removed from the notes, when present, using regular expressions. The dataset was divided into 70%/15%/15% subsets for training, validation and testing. Baseline identification was calculated by a simple lexical matching using the term “synkope”. We evaluated two linear classifiers, a Support Vector Machine (SVM) and a Linear Regression (LR) model, with a term frequency–inverse document frequency vectorizer, using a bag-of-words approach. In addition, we evaluated a simple convolutional neural network (CNN) consisting of a convolutional layer concatenating filter sizes of 3–5, max pooling and a dropout of 0.5 with randomly initialised word embeddings of 300 dimensions.
Results
Even a baseline regular expression model achieved a sensitivity of 78% and a specificity of 91% when classifying admission notes as belonging to the syncope class or not. The SVM model and the LR model achieved a sensitivity of 91% and 89%, respectively, and a specificity of 89% and 91%. The CNN model had a sensitivity of 95% and a specificity of 84%.
Conclusion
With a limited non-English dataset, common NLP and machine learning approaches were able to achieve approximately 90–95% sensitivity for the identification of admission notes related to syncope. Linear classifiers outperformed a CNN model in terms of specificity, as expected in this small dataset. The study demonstrates the feasibility of training document classifiers based on diagnostic codes in order to detect important clinical events.
ROC curves for SVM and LR models
Funding Acknowledgement
Type of funding source: Public grant(s) – National budget only. Main funding source(s): The Research Council of Norway
Natural Language Processing as an Emerging Tool to Detect Late-Life Depression
