Inferring insulin regimen from clinical notes : using natural language processing techniques to extract data from free text records

Abstract

Insulin Regimen refers to instructions prescribed by a clinician indicating the kind of insulin to take (long-acting, short-acting, etc), how much insulin to take (dosage) and how often to take it (frequency). Determining the daily insulin regimen for diabetic patients is more of an art than a science. Clinicians who care for diabetic patients carefully assess the patient's blood glucose levels, medical history and symptoms before prescribing insulin medication. The challenge for clinicians is often in accessing the historical insulin regimen prescribed to patients, which is hidden in unstructured clinical notes. The reason that is a problem is that the individual clinician is unable to draw on the wisdom that might exist in collective experience. Additionally, having access to a patient's historical insulin regimen can help identify patient groups with distinct insulin regimen patterns, analyze total and average daily insulin consumption of different patient groups, discover patient groups showing variation in their insulin regimen, etc. In this thesis, we treat insulin regimen extraction from clinical notes as an information extraction problem and explore machine learning methods focused on extracting this information from prescription lists available in outpatient clinical notes. We explore two n-gram models - Logistic Regression and Conditional Random Field and analyze their performance. We also explore models using contextual word representations from the domain specific pretrained language models, character level embeddings and auxillary features constructed from external knowledge sources and analyze their performance. We find that our final Multi Layer Perceptron method using contextual word representations gives a micro averaged F1 score of 0.98 and is able to detect patterns that go undetected by n-gram models. We then apply a rule based post processing system to convert the extracted insulin regimen into a normalized timeseries format. We analyze the extracted insulin regimen information and find that, in most cases, prescription lists in clinical notes contain an accurate account of the current insulin regimen prescribed to patients. However, supporting insulin regimen information such as patient specific glycemic targets, basal-bolus insulin ratio, etc are available only in the narrative text in clinical notes. We also examine the data to find patient samples with interesting insulin regimen patterns such as those changing from a long-short to combined insulin regimen and vice versa from our extracted dataset.