Bridging Data Science Approaches to Improve the Health and Care of Children of Incarcerated Parents

Abstract

Research ObjectiveMass incarceration has had an undeniable toll on childhood poverty and inequality, however, little is known about the consequences on pediatric health. Over 5.7 million children, or 1 in every 14 children, have experienced a parent in prison or jail at some time point. Because families in poverty, families of color, and families in rural areas are more likely to be affected, parental incarceration is viewed as a key driver of racial and economic disparities in health. Yet, we fail to routinely screen for this exposure, in part, due to stigma of incarceration, and the limited resources and time in the clinical setting. Our objective was to 1) describe the health of youth who have a documented correctional keyword in their chart 2) develop a machine learning model to locate children of incarcerated parents (COIP).Study DesignA descriptive study was first conducted using electronic health record data of 2.3 million youth (ages 0–21 years) who received care in a large Midwestern hospital‐based institution from February 2006–2020. We employed a correctional‐related keyword search (e.g. jail, prison, probation, parole) to locate youth with probable personal or family history involvement. Health characteristics were measured as clinician diagnostic codes. We then conducted a manual chart annotation of 10,000 random cases using correctional related keyword search terms to annotate cases as 1 (“have or had a parent in jail/prison”) or 0 (“false positive”). This resulted in a dataset suitable to train BERT, a state‐of‐the‐art machine learning model which is unique in its ability to pick up context within and across sentences.Population StudiedElectronic health record data of 2.3 million youth (ages 0–21 years) who received care in a large Midwestern hospital‐based institution.Principal FindingsTwo percent of the total pediatric population had a correctional keyword in the medical chart (N = 51,855). This 2% made up 66% of all patients with cannabis‐related diagnoses, 52% of all patients with trauma‐related diagnoses, 48% of all stress‐related diagnoses, 38% of all patients with psychotic disorder diagnoses, and 33% of all suicidal‐related disorders within this institution's electronic health record database – among other highly concerning findings. The development of the machine learning model to specifically locate a child who has been exposed to parental incarceration was feasible.ConclusionsWe captured an alarming health profile that warrants further investigation and validation methods to better address the gaps in our clinical understanding of youth with personal or family history involvement with the correctional system. There is also a need for robust language model testing based on machine learning in order to correctly identify evidence of COIP in the electronic health record.Implications for Policy or PracticeWe can do better in identifying, and supporting families affected by the correctional system. This novel cohort identification method may be able to fulfill the gaps in the sciences related to COIP or other types of family and personal involvement with the justice system. Doing so, could inform intervention development, and effective policy creation to improve the care and health of those exposed.Primary Funding SourceNational Center for Advancing Translational Sciences.