Documentation of the patient’s smoking status in common chronic diseases – analysis of medical narrative reports using the ULMFiT based text classification

INTRODUCTION: Smoking cessation is essential part of a successful treatment in many chronic diseases. Our aim was to analyse how actively clinicians discuss and document patients’ smoking status into electronic health records (EHR) and deliver smoking cessation assistance. METHODS: We analysed the results using a combination of rule and deep learning-based algorithms. Narrative reports of all adult patients, whose treatment started between years 2010 and 2016 for one of seven common chronic diseases, were followed for two years. Smoking related sentences were first extracted with a rule-based algorithm. Subsequently, pre-trained ULMFiT-based algorithm classified each patient’s smoking status as a current smoker, ex-smoker, or never smoker. A rule-based algorithm was then again used to analyse the physician-patient discussions on smoking cessation among current smokers. RESULTS: A total of 35,650 patients were studied. Of all patients, 60% were found to have a smoking status in EHR and the documentation improved over time. Smoking status was documented more actively among COPD (86%) and sleep apnoea (83%) patients compared to patients with asthma, type 1&2 diabetes, cerebral infarction and ischemic heart disease (range 44-61%). Of the current smokers (N=7,105), 49% had discussed smoking cessation with their physician. The performance of ULMFiT-based classifier was good with F-scores 79-92. CONCLUSION: Ee found that smoking status was documented in 60% of patients with chronic disease and that the clinician had discussed smoking cessation in 49% of patients who were current smokers. ULMFiT-based classifier showed good/excellent performance and allowed us to efficiently study a large number of patients’ medical narratives.