Targeted Neural Text Summarization of Electronic Medical Records to Improve Imaging Diagnostics

点击翻译按钮获取中文摘要

Project Summary Targeted Neural Text Summarization of Electronic Medical Records to Improve Imaging Diagnosis Electronic health records (EHRs) contain a wealth of patient information that might inform diagnostic and therapeutic decision-making. However, much of this information is unstructured (i.e., free-text). This makes it difficult to find the few relevant notes that might inform a given decision amongst lengthy patient records, in turn rendering key information buried within EHR practically inaccessible to domain experts operating under time constraints. Consequently, clinical decisions are often made without the benefit of all available data. We propose to design, train, and deploy novel natural language processing (NLP) models that provide extractive summaries of the free-text data within EHR conditioned on particular queries; the intent is for such models to aid diagnosis and decision-making. We also propose to use these models to try and counteract the cognitive biases that domain experts bring to clinical practice. We focus specifically on the important and illustrative area of radiology, although the approach will generalize to other specialties. Radiologists performing imaging diagnosis do not have adequate time to carefully read through patient histories stored within EHR; they must instead make do with limited background information when interpreting imaging. We will build on our preliminary on models that summarize textual evidence extracted from EHR that might support particular hypothesized diagnoses. We envision an interactive system in which this model is used by the radiologist to surface textual evidence that supports different potential conditions that might be suggested by the imaging. Radiologists (and other domain experts) rely on heuristics — type 1 thinking — when making decisions under time constraints. This results in various cognitive biases influencing diagnoses, and these have been shown to be the source of a significant fraction of diagnostic errors in radiology. We propose a novel secondary use of the NLP models to be developed for this project as a means of counteracting these cognitive biases. Specifically, once the radiologist has indicated an initial potential diagnosis via a natural language query, we will automatically present a few alternative plausible diagnosis and summaries of the extracted evidence supporting these (alongside the summary of evidence relevant to the initial query). These alternative diagnoses will be gleaned from gamuts or published lists of differential diagnoses, and we will re-rank them in order of their predicted probability for the current patient according a trained machine learning model. We will evaluate the proposed models in practice at Brigham and Women's Hospital, and assess the degree to which integrating automatically generated summaries actually affects clinical decision-making at point of care.