Automating Assessment of Contextualization of Care During the Clinical Encounter

点击翻译按钮获取中文摘要

Background: Large scale studies have demonstrated that when patients struggle with life challenges that complicate their care (e.g., a lack of social support, competing responsibilities, or financial hardships), health care providers can improve health care outcomes and lower costs if they attempt to identify these “contextual factors” and address them in their care plan – a process termed “contextualizing care.” These studies utilize a method of data analysis called “Content Coding for Contextualization of Care” (4C). 4C is a labor-intensive process that requires human coders listen to audio recordings from a sample of medical encounters, extract data from each patient’s medical record, and then track whether identified contextual factors have been addressed. It has enormous social value and commercial potential because it accurately identifies care plans that mitigate social needs and avoid unnecessary care. To date, sharing 4C data with health systems has led to significant improvements in care and a reduction in rates of hospitalization. The manual 4C coding process, however, is time consuming and unscalable. The automation of 4C coding utilizing natural language processing (NLP) would enable rapid scaling. Objective: Build a prototype system that performs automated 4C coding of transcribed audio-recorded medical encounters, and assess its accuracy at classifying whether care planning is contextualized, utilizing a test dataset with human 4C coding as a gold-standard. Method: We propose an iterative development and validation process, leveraging an existing dataset of over 400 manually 4C coded transcripts from physician-patient medical encounters. Starting with 300 transcripts previously coded by our team and coding guidelines from the 4C training manual, we will first develop automated techniques for extracting text features reflective of nuances in linguistic content and discourse structure that disentangle contextualized care from contextual error, in turn facilitating development of candidate classification models that emulate human 4C coding decisions. We will then apply the models to the remaining transcripts to predict transcript- and utterance-level codes, comparing these codes with the human-labeled gold standard to establish feasibility, analyze performance, and assess the models’ performance when generalized to new clinical encounters. Impact: Health care systems are under financial pressure to control costs through a reduction in both preventable hospitalizations and overuse and misuse of medical services. This phase 1 STTR will establish the feasibility and technical merit of automated 4C coding to provide a low cost, scalable strategy for accurately measuring and facilitating clinical performance that enhances value-based care. Such technology is especially timely as audio recording visits is increasingly common as virtual scribes work remotely to document medical visits, and audio recordings are provided as an information aid to patients.