A natural language processing approach to identify social risk factors from neonatal intensive care unit patients’ (NICU) electronic health records

Adverse social drivers/determinants of health (SDOH) such as food, housing, and income insecurity (hereby referred to as social risk factors) can have significant impact on an infant’s health and may exacerbate health inequities. To reduce health burdens and improve clinical decision-making, it is necessary to understand how

social risk factors are associated with hospital readmissions among infants in the neonatal intensive care unit (NICU). However, measuring social risk factors is challenging, as SDOH are not well documented in the electronic health record (EHR), and current EHR platforms lack screening tools to identify social risk factors in

NICU clinical settings. Patient clinical notes contain a wealth of SDOH information, but researchers face the challenge of manually reviewing and extracting this information, which is time-consuming. The objective of this proposal is to: 1) leverage natural language processing to identify and extract social risk factors

from unstructured clinical note EHR data of NICU patients and 2) understand how unmet social risk factors predict hospital readmissions among infants discharged from the NICU using both structured and unstructured EHR data. To achieve this objective, this study will fine-tune developed Bidirectional

Encoder Representations from Transformers (BERT) models on NICU clinical notes to improve their ability to recognize and extract social risk factors (Aim 1). Investigators will be able to better identify and classify infants with food, housing, and income insecurity, and understand how social risk factors are characterized in NICU

clinical notes. The performance, validation, and generalizability of these models will be evaluated with standard metrics. Additionally, this study will assess how social risk factors are associated with 30-day hospital readmission among infants discharged from the NICU (Aim 2). Our central hypothesis is we will be able to

uncover additional information related to a patient’s experience with social risk factors, and these risk factors will be associated with higher rates of 30-day hospital readmission. We will explore how associational effect estimates change using solely structured data (diagnosis codes) compared to using both structured and

unstructured data. This project will advance steps to streamline and automate the extraction of valuable SDOH data from clinical notes and guide clinical decision-making in NICU care. The long-term goal of this project is to harness the power of advanced data science methods to optimize patientcare and management for unmet

social needs and to improve health equity thus reducing health utilization outcomes. This research aligns well with NIMHD’s Division of Clinical and Health Service Research. The proposed training, guided by an expert mentorship team, will enrich the applicant’s knowledge of and skills in clinical informatics for health disparities,

data science for social epidemiology, and natural language processing for classification tasks. The content expertise, research competency, and training in quantitative methods the applicant will receive will prepare her well to improve scientific knowledge and clinical practice in her career as an independent researcher.