Clinical Data Extraction and Normalization of Cyrillic Electronic Health Records Via Deep-Learning Natural Language Processing

Abstract

PURPOSEA substantial portion of medical data is unstructured. Extracting data from unstructured text presents a barrier to advancing clinical research and improving patient care. In addition, ongoing studies have been focused predominately on the English language, whereas inflected languages with non-Latin alphabets (such as Slavic languages with a Cyrillic alphabet) present numerous linguistic challenges. We developed deep-learning–based natural language processing algorithms for automatically extracting biomarker status of patients with breast cancer from three oncology centers in Bulgaria.METHODSWe used dual embeddings for English and Bulgarian languages, encoding both syntactic and polarity information for the words. The embeddings were subsequently aligned so that they were in the same vector space. The embeddings were used as input to convolutional or recurrent neural networks to derive the biomarker status of estrogen receptor, progesterone receptor, and human epidermal growth factor receptor 2.RESULTSWe showed that we can resolve ambiguity in highly variable medical text containing both Latin and Cyrillic text. Final models incorporating both English and Bulgarian syntax and polarity embeddings achieved F1 scores of 0.90 or higher for all estrogen receptor, progesterone receptor, and human epidermal growth factor receptor 2 biomarkers. The models were robust against human errors originally found in the training set. In addition, such models can be extended for analyzing text containing words not seen during training.CONCLUSIONBy using several techniques that incorporate dual-word embeddings encoding syntactic and polarity information in two languages followed by deep neural network architectures, we show that researchers can extract and normalize parameters within medical data. The principles described here can be used to analyze Cyrillic or Latin mixed medical text and extract other parameters.

Machine Learning-Based Extraction of Breast Cancer Receptor Status From Bilingual Free-Text Pathology Reports

“…This study was designed to evaluate the possibility of automatically extracting the status of the 3 main breast cancer biomarkers (ER, PR, and HER2) from the contents of pathology reports written in two different languages, and coming from 82 different providers, using conventional machine learning models. After testing different classifiers, the best performing ones achieved macro-averaged F 1 scores ranging from 0.89 to 0.92 on the held-out test sets, which is on par with best efforts in the literature (6,7,11,12). The reported F 1 scores in the literature range between 0.87 and 1, but use only three possible labels for HER2, whereas five are used in the present work.…”

“…The extraction is performed at a national level, involving 82 different data providers (all Belgian laboratories for pathological anatomy). This problem has not been addressed in the previously mentioned studies, which focused on a limited number of data sources (6,(11)(12)(13).…”

“…Within the context of these activities, having the breast cancer receptor status at its disposal would undoubtedly be of added value. In breast cancer, estrogen receptor (ER), progesterone receptor (PR), and Erb-b2 receptor tyrosine kinase 2 (ERBB2, previously named Human Epidermal Growth Factor 2 or HER2 or HER-2/neu 2 ) are biomarkers known to be related to tumor growth and prognosis, and assessing their expression is necessary to define therapeutic management (3)(4)(5)(6)(7). Currently this information is not available in a structured form at the BCR.…”

Status and Recommendations of Technological and Data-Driven Innovations in Cancer Care: Focus Group Study

“…In terms of data interpretation and translation, we recommend the use of tools to extract and represent the medical substrate by synthesizing only relevant aspects in a declarative way. ML techniques (ie, deep learning—recurrent networks with word embeddings and distributed representations) can handle very large and sparse data (eg, device data may only be available for a small subset of individuals) to capture the sequential character of the data and are suitable for modeling context dependencies in inputs [ 57 ]. Such systems, which incorporate word embeddings encoding syntactic and polarity information in the language followed by deep neural network architectures, are already used to extract and normalize parameters within oncology care data.…”

Triple negative breast cancer in Bulgaria: epidemiological data and treatment patterns based on real world evidence and patient registries

“…In the era of innovations in digital technologies and value-based patient care substantial part of the medical information is still unstructured. Extraction and systematization of information from medical records are of a great significance for improvement of diagnosis, treatment, survival prediction, resource allocation and decision making [10]. Hence, in recent years there is an increased interest towards the term 'big data' , its interpretation and potential use in health economics and outcomes research (HEOR) [11].…”