Using Natural Language Processing to Identify and Classify Incidental Findings

Abstract

Over the last few years there has been an explosion of deep learning research leading to quick development of very powerful algorithms that have found their way into every industry. There are a few major reasons why deep learning is particularly attractive for algorithm development. The algorithms can complete complex tasks with surprisingly high accuracy. Deep learning models, unlike machine learning models, automatically extract features that are important for the algorithms task. The deep learning community has established and open-sourced high quality feature extractors and classifier architectures with weights included for most data domains. Lung cancer is a major cause of morbidity and mortality in the US and globally. Lung cancers are conditions that include a silent phase during which intervention is highly effective, but patients are asymptomatic so there is no indication for an exam. It is common that patients with symptoms unrelated to a forming cancer get imaging to work-up their current illness. During these times an incidental nodule in the lungs may be captured and a radiologist may recommend a follow-up, but there are few mechanisms in place to ensure that these patients have their follow-up completed. As a point of quality control, NM would like to maximize the likelihood that a patient with an incidentally noted lung nodule with follow-up recommendations will receive appropriate follow-up. In order to meet this goal, we have developed an EHR ready artificial intelligence pipeline that identifies reports containing text suggesting a lung nodule requiring follow-up. We find that upon retrospective review, we are able to use machine learning to capture reports containing lung nodules requiring follow-up with a sensitivity of 87% and specificity of 87%.

Files

Additional details