Natural language processing for disease phenotyping in UK primary care records for research: a pilot study in myocardial infarction and death

Abstract

BackgroundFree text in electronic health records (EHR) may contain additional phenotypic information beyond structured (coded) information. For major health events – heart attack and death – there is a lack of studies evaluating the extent to which free text in the primary care record might add information. Our objectives were to describe the contribution of free text in primary care to the recording of information about myocardial infarction (MI), including subtype, left ventricular function, laboratory results and symptoms; and recording of cause of death. We used the CALIBER EHR research platform which contains primary care data from the Clinical Practice Research Datalink (CPRD) linked to hospital admission data, the MINAP registry of acute coronary syndromes and the death registry. In CALIBER we randomly selected 2000 patients with MI and 1800 deaths. We implemented a rule-based natural language engine, the Freetext Matching Algorithm, on site at CPRD to analyse free text in the primary care record without raw data being released to researchers. We analysed text recorded within 90 days before or 90 days after the MI, and on or after the date of death.ResultsWe extracted 10,927 diagnoses, 3658 test results, 3313 statements of negation, and 850 suspected diagnoses from the myocardial infarction patients. Inclusion of free text increased the recorded proportion of patients with chest pain in the week prior to MI from 19 to 27%, and differentiated between MI subtypes in a quarter more patients than structured data alone. Cause of death was incompletely recorded in primary care; in 36% the cause was in coded data and in 21% it was in free text. Only 47% of patients had exactly the same cause of death in primary care and the death registry, but this did not differ between coded and free text causes of death.ConclusionsAmong patients who suffer MI or die, unstructured free text in primary care records contains much information that is potentially useful for research such as symptoms, investigation results and specific diagnoses. Access to large scale unstructured data in electronic health records (millions of patients) might yield important insights.

The Potential of Research Drawing on Clinical Free Text to Bring Benefits to Patients in the United Kingdom: A Systematic Review of the Literature

“…The simplest form of information extraction was manual review of free text ( 40 – 45 ) or a keyword search for relevant information, which was then modified by an algorithm ( 46 ) or manual review ( 47 – 49 ) to check for negation and uncertainty. Shah et al described a bespoke algorithm for converting general practice text data into categorical data for analysis ( 50 ). Papers drawn from the SLAM CRIS database used a bespoke set of data extraction techniques, supplied in the CRIS system using the Generalized Architecture for Text Engineering (GATE) ( 51 ) which allows integration of a range of NLP algorithms for specific purposes, such as identification of medications or Mini-Mental State Examination (MMSE) scores for dementia.…”

“…Medication information was often extracted from free text, particularly in studies using the SLaM BRC CRIS case register ( 57 , 60 , 63 , 64 , 66 – 73 ). Also extracted from free text were disease symptoms and drug reactions ( 46 , 47 , 50 , 60 , 67 , 72 , 74 – 76 ); test scores, such as for the MMSE ( 58 , 59 , 61 , 63 , 64 , 77 , 78 ), and angiogram results ( 50 ); treatments such as cognitive behavioral therapy (CBT) ( 60 , 79 ); substance use behaviors such as cannabis ( 49 , 72 , 80 ), alcohol ( 43 , 44 , 49 ) or smoking status ( 81 ); housing status ( 45 ); and information on symptom severity and functional status ( 61 , 73 ).…”

Predicting need for hospital admission in patients with traumatic brain injury or skull fractures identified on CT imaging: a machine learning approach

“…Such clinical scenarios include classification tasks related to medical imaging 18 or the natural language processing of freetext health records. 19 Such research should be reported, transparently and according to consistent reporting standards, such as those that build on the TRIPOD guidelines for prognostic studies. 20 Our machine learning models would select patients for discharge with around a 1 in 26 chance of subsequently deteriorating.…”

Estimating the population health burden of musculoskeletal conditions using primary care electronic health records

“…Lifestyle information, in particular, can often be missing from these records, but completeness has been improving over recent years. Performance of models could be improved by including information from the unstructured free text within the EHR [ 29 ] but access to this is increasingly difficult for researchers in the UK due to information governance restrictions. The prediction models have been derived using retrospective data and are limited in their application at the individual level to identify those at high risk.…”