Building a tobacco user registry by extracting multiple smoking behaviors from clinical notes

Abstract

Background Usage of structured fields in Electronic Health Records (EHRs) to ascertain smoking history is important but fails in capturing the nuances of smoking behaviors. Knowledge of smoking behaviors, such as pack year history and most recent cessation date, allows care providers to select the best care plan for patients at risk of smoking attributable diseases. Methods We developed and evaluated a health informatics pipeline for identifying complete smoking history from clinical notes in EHRs. We utilized 758 patient-visit notes (from visits between 03/28/2016 and 04/04/2016) from our local EHR in addition to a public dataset of 502 clinical notes from the 2006 i2b2 Challenge to assess the performance of this pipeline. We used a machine-learning classifier to extract smoking status and a comprehensive set of text processing regular expressions to extract pack years and cessation date information from these clinical notes. Results We identified smoking status with an F1 score of 0.90 on both the i2b2 and local data sets. Regular expression identification of pack year history in the local test set was 91.7% sensitive and 95.2% specific, but due to variable context the pack year extraction was incomplete in 25% of cases, extracting packs per day or years smoked only. Regular expression identification of cessation date was 63.2% sensitive and 94.6% specific. Conclusions Our work indicates that the development of an EHR-based Smokers’ Registry containing information relating to smoking behaviors, not just status, from free-text clinical notes using an informatics pipeline is feasible. This pipeline is capable of functioning in external EHRs, reducing the amount of time and money needed at the institute-level to create a Smokers’ Registry for improved identification of patient risk and eligibility for preventative and early detection services. Electronic supplementary material The online version of this article (10.1186/s12911-019-0863-3) contains supplementary material, which is available to authorized users.

Impact of deep learning-determined smoking status on mortality of cancer patients: never too late to quit

“…Artificial intelligence and deep learning applications are making their way into clinical practice, but their efficacy is not yet proven in prospective trial settings. 16 The current study utilized more advanced deep learning compared to previous reports 18 , 19 , 20 in a retrospective proof-of-principle setting, where we successfully extracted a large amount of smoking data in a matter of days. Two language models were tested with good specificity (88%-98%), comparable to the previous results in English language.…”

“…24 The learned knowledge was then transferred to a training classifier, by randomly picking 5000 tobacco smoking-related sample phrases and sentences from the medical narrative archive of our hospital, using the Finnish word-stem ‘tupak’ equivalent to the English word-stem ‘smok’. 19 These sample phrases were manually labeled into three classes (never, former, or current smoker). ULMFiT- and BERT-based classification models were then trained on this data to produce smoking phrase classifiers.…”

“…SPSS version 26 (IBM, Armonk, NY) was used. Sensitivity and specificity analyses were calculated with 2 × 2 contingency tables separately for never, former, and persistent smokers, excluding patients with missing smoking status, 19 using the Python ‘sklearn’ package.…”

“… 16 , 17 This includes universal language modeling of medical narratives, wherein real-life practice health data, such as smoking status, are often presented in an unstructured format. 18 First reports of using language modeling to define an individual's smoking status appeared encouraging both from electronic health records (EHRs) 18 , 19 and from user-generated content on a smoking cessation support website. 20 …”

Oral Human Papillomavirus: a multisite infection

“…OSCG were subjected to biopsy to obtain diagnostic certainty. The data on tobacco consumption was made according the number of pack year smoked; a subject that smoked fifteen or more pack years in twenty years was considered a smoker (13). One alcohol unit a day (one drink) was considered as regular alcohol consumption (14).…”