Considerations for development of child abuse and neglect phenotype with implications for reduction of racial bias: a qualitative study

Abstract

Objective

The study provides considerations for generating a phenotype of child abuse and neglect in Emergency Departments (ED) using secondary data from electronic health records (EHR). Implications will be provided for racial bias reduction and the development of further decision support tools to assist in identifying child abuse and neglect.

Materials and Methods

We conducted a qualitative study using in-depth interviews with 20 pediatric clinicians working in a single pediatric ED to gain insights about generating an EHR-based phenotype to identify children at risk for abuse and neglect.

Results

Three central themes emerged from the interviews: (1) Challenges in diagnosing child abuse and neglect, (2) Health Discipline Differences in Documentation Styles in EHR, and (3) Identification of potential racial bias through documentation.

Discussion

Our findings highlight important considerations for generating a phenotype for child abuse and neglect using EHR data. First, information-related challenges include lack of proper previous visit history due to limited information exchanges and scattered documentation within EHRs. Second, there are differences in documentation styles by health disciplines, and clinicians tend to document abuse in different document types within EHRs. Finally, documentation can help identify potential racial bias in suspicion of child abuse and neglect by revealing potential discrepancies in quality of care, and in the language used to document abuse and neglect.

Conclusions

Our findings highlight challenges in building an EHR-based risk phenotype for child abuse and neglect. Further research is needed to validate these findings and integrate them into creation of an EHR-based risk phenotype.

BACKGROUND AND SIGNIFICANCE

Child abuse and neglect are defined as any action (physical, emotional, and/or sexual) by a primary caregiver that results in harm, potential harm, or threat of harm to a child.¹ Child abuse and neglect have reached epidemic proportions, with more than 700,000 victims of abuse annually in the United States.² In clinical practice, it is challenging to identify potential child abuse and neglect cases.^3,4 First, little is known about factors that are associated with increased risk of child abuse and neglect.⁵ In addition, recent studies have highlighted that racial bias begets oversuspicion of neglect and abuse within Black and Latinx communities.⁶ In 2019, approximately 45% of identified child abuse victims in the United States were Black and Latinx.⁷ Another study that used state-wide data from California, found that about 50% Black and Native American children were investigated during childhood for suspicion of child abuse and neglect.⁸ Large-scale adoption of electronic health records (EHR) in clinical settings offers a new avenue for addressing child abuse and neglect.^9,10 Despite the potential for EHR data to improve child abuse and neglect detection abilities, the analysis of EHR data is complex and technically challenging due to scattered and fragmented documentation.¹¹ Furthermore, the concern for existing racial bias in EHRs creates unique challenges in identifying a risk phenotype that does not further propagate existing racial biases^12,13 within the detection and reporting of child abuse and neglect. For example, recent research that used EHR data to develop automated identification of risk for child abuse and neglect did not address racial bias.⁹

This exploratory research project aims to develop an artificial intelligence (AI)-based clinical decision support system to identify potential child abuse and neglect by processing pediatric emergency department (ED) EHR data. Our first step was to interview clinicians (pediatricians, nurses, social workers, and other medical staff) to gain their insights regarding risk factors, assessment, documentation, and challenges of diagnosing child abuse and neglect in EDs.

OBJECTIVE

This article will discuss considerations for generating a phenotype of child abuse and neglect in EDs using secondary data from EHR. Implications will be provided for racial bias reduction and the development of further clinical decision support tools to assist in identifying child abuse and neglect.

MATERIALS AND METHODS

We conducted a qualitative study through in-depth interviews with 20 pediatric clinicians working in a single pediatric, tertiary-care ED to gain insights about generating an EHR-based phenotype to identify children at risk for abuse and neglect.

Participants

All study participants were affiliated with a single pediatric ED in the Northeast region of the United States. Clinicians were recruited using both a snowball and purposive sampling strategy. Snowball recruitment methods included asking interviewees to identify potential colleagues willing to participate. Purposive sampling strategies consisted of formal and informal invitations to participate in the research, such as emails, phone calls, presentations during departmental meetings, and established relationships with clinicians. Inclusion criteria encompassed: (1) pediatric clinicians who provided clinical care for patients suffering from abuse and neglect within the pediatric ED, (2) clinicians who can share insights regarding medical documentation concerning child abuse and neglect, and (3) clinicians who can provide their perspectives and experiences on providing care to marginalized communities treated within the pediatric ED.

Procedure

One-to-one interviews were conducted between August and December 2020. The sample size was determined by theoretical saturation; interviews continued until no new themes were identified during ongoing review of transcripts.¹⁴ Interviews were conducted by AL, a social work scientist and trained qualitative researcher. AL is a white Ashkenazi Jew and a third-generation Holocaust survivor. His previous experience working with vulnerable populations and his current research surrounding child abuse and neglect and racial bias are informed by a commitment to promote social justice and equity by developing innovative intervention tools that support marginalized communities. All 20 semistructured interviews were conducted via Zoom due to Covid-19 health protocols. Online interviewing enabled flexible scheduling opportunities and locations, thus increasing availability and response (80% approved response rate). Interviews were between 45 min and 2 h long. All interviews were audio-recorded and transcribed using Zoom transcription before analysis. Study participants received a $100 gift card for their participation in the research.

The interview questions were designed to elicit insights on racial bias, EHR documentation, and personal clinical experience with child abuse and neglect. We used the following questions to guide the semistructured interviews: (1) What are clinicians’ challenges in diagnosing child abuse and neglect? (2) How do clinicians document child abuse and neglect in the EHR? (3) How does racial bias impact the diagnosis of child abuse and neglect?

Data analysis

The majority of the research team is trained in qualitative research (AL, SS, NA, MT, KC, DP). We applied thematic analysis to identify central themes and categories^14–16 via the following 3 phases. In phase 1, 3 members of the research team employed open coding to achieve analyst triangulation by reading the interviews several times to become familiar with the content and identify central themes and categories for developing an initial codebook. This method provides a check on selective perceptions and blind spots while doing interpretive analysis.¹⁷ Next, using Dedoose¹⁸ qualitative software, the researchers AL, SS, and NA first coded 3 identical transcripts individually to explore different meanings and patterns within the data, identify initial areas of disagreement to discuss and agree upon within the larger group, refine the codebook, and establish a final code scheme. In this phase, we developed codes such as “bias in diagnosis of abuse and neglect.”

In phase 2, the researchers continued with the analysis of the remaining interviews and compared codes while simultaneously grouping central codes and sorting data into initial categories.^14–16 Categories were compared, aligned, and gathered into central themes that provide meaning and highlight overarching patterns in the data. As the analysis continued, categories were added, removed, and changed based on the emerging content. For example, first, we created and used the codes “Different styles of documentation” and “key words to describe abuse and neglect” to label descriptions of how various clinical disciplines use different styles and language in clinical notes when describing child abuse and neglect. Second, these codes were aggregated into categories such as “Examples of language used in clinical notes describing child abuse and neglect.” Third, we grouped this category under the overarching theme “Different documentation styles in EHR,” which highlights essential differences in clinical documentation concerning abuse and neglect.

In phase 3, the final list of central themes was discussed, refined, and finalized by the research team at large via consensus.¹⁶ We divided existing categories to further specify the challenges clinicians face in diagnosing child abuse and neglect. For example, a category named “challenges of identifying abuse” was divided into 2 new categories: “difficulties in diagnosis due to ambiguity concerning the symptoms/signs of potential abuse/neglect,” and “difficulties in diagnosis due to lack of continuity of care.”

Ethics

This study received institutional review board approval from Columbia University. Prior to conducting the interviews, we received informed consent from all clinicians. During the consent process, we presented the aim of this study, both in writing and verbally, including the right to refuse to participate or to end the interview.

RESULTS

The clinicians interviewed reflect an interdisciplinary and experienced team whom patients may encounter in the ED and whose EHR documentation we analyze for the development of an AI-based clinical decision support system to identify potential child abuse and neglect. The clinicians included social workers (n = 3, 100% of social workers working in the ED at that time), nurses (n = 5), pediatricians (n = 10), a registered respiratory therapist (n = 1), and a physician’s assistant (n = 1) who worked in the pediatric ED, where the study was conducted, for at least the past year (range, 1–30, mean= 10.15 years). The average age of the clinicians was 41 (range, 27–66) years old. The clinicians interviewed were also racially and ethnically diverse, with 40% of the interviewees (n = 8) identified as Black or Latinx, 15% (n = 3) identified as Asian, and 45% identifying as white.

Three final central themes emerged from the interviews, including:

1. Identifying and diagnosing child abuse and neglect in the ED is challenging due to work conditions, symptom ambiguity, and scattered and fragmented information in the EHR.

2. Differences in documentation style exist between health disciplines, including various documentation locations, styles, and phrasing of clinicians working in the ED.

3. EHR data can reveal potential racial bias in quality of care and through differences in vocabulary used in medical documentation.

Theme 1: challenges in diagnosing child abuse and neglect

A vast majority (85%) of clinicians described specific challenges in diagnosing child abuse and neglect in the ED. The challenges were in part due to work conditions, such as time of shift and volume of patients, lack of detailed information regarding patients, and the difficulty of diagnosing abuse and neglect through brief examinations and conversations with families (Table  1). Clinicians discussed the importance of both time and continuity of care for appropriately identifying child abuse and neglect (quote 1 [Q1]). They explained that meeting patients irregularly can decrease clinicians’ ability to understand and detect patterns of abuse (Q2).

Table 1.

Challenges in diagnosing child abuse and neglect

Difficulties in diagnosis due to lack of continuity of care Quote 1 [Q]1: “I think it's hard because we only see them for an hour and that's it, so I think it's hard as emergency room providers in being able to identify it because we really only have such limited amount of time with them (patients). I think that's what makes it most difficult that there is no continuity of care.” (Physician Assistant) Q2: “A big problem—when it comes to identifying child abuse in kids is that my understanding is that usually there's patterns of injury, and we're not necessarily seeing the same patient again and again and again, so whereas a pediatrician might see the same kid multiple times. (.…) I think it's much easier to not pick up on it when it's like a quick ER visit.” (Nurse)

Difficulties in diagnosis due to patient volume or admission timing Q3: “So I think that the sheer volume of the patients that come in… it's overwhelming, I mean there are times where we're just barely getting through the day. (. . . .) The challenges of recognizing child neglect. Some of the challenges again, like I said, are the time constraints. If a physician has 20 kids, and you only have a few minutes to spend with those 20 kids, but you have 5 that are critically in need of attention, you're going to spend your time with those 5.” (RRT- Registered Respiratory Therapist) Q4: “When there's no (pediatric) social worker (at night). In drastic cases, sometimes the social worker from the adult ER will help you. But then that's a little bit harder because they're so busy in the adult ER, so you're waiting. We have an excellent pediatric social work team. So, you almost kind of want to wait until morning because you know that they'll be on top of everything. And they make sure everything is done properly.” (Physician)

Difficulties in diagnosis due to ambiguity concerning the symptoms/signs of potential abuse/neglect Q5: “Sexual abuse you know probably 99% - a very high percentage of sexual abuse has nothing on physical exam so (. . . .) I usually will say in my note, ‘physical exam appears normal; however, a normal physical exam does not rule out sexual abuse as the majority of cases of sexual abuse have normal physical exam findings.” (Physician) Q6: “He was placed in foster care [patient] (. . . .) and the birth mother brought him in and said that, ‘my child lost a lot of weight; he's not eating. (. . . .) He's always—he said he's always in pain,’ so in my mind it could be child abuse. It could be many things. You know, it could be an illness, so we brought him to the acute side of the E.R., and we ran out a bunch of tests, and it turns out that he has new onset cancer.” (Nurse)

Subtle information Q7: “Neglect is a little harder because we can't see whether they're being fed at home. It's not physical, but more like emotional as well sometimes— ‘did you neglect the child emotionally?’, but I think that's just a challenge, the fact that it's not so obvious. You have to really be paying attention to the clues around you.” (Nurse) Q8: “Neglect cases are a harder thing to prove because I've actually had a kid saying, ‘mom isn't giving me food,’, and ‘not providing for me,’ which is technically neglect, and then the mom is like, ‘no, I give him. He just doesn't like what I give him,’ so you have to talk to both parents and find out what's actually happening because sometimes with neglect you hear one thing from the kid and a different thing from the parent, or there's a newborn in the house, so they feel neglected but it's not actual neglect.” (Social Work) Q9: “First of all, sometimes it's not identified in triage because the mother doesn't come right off and say, ‘I'm here because my boyfriend hit my daughter.’ Sometimes it could just be that my daughter has a stomach ache that's not going away, and then it's not identified.” (Physician)

Scattered and fragmented or lack of information in EHR Q10: “You might look for a social work note if you're a little bit concerned or you pull up their chart and you happen to see like a gazillion different x-rays over a year like that might be concerning, but I think that unless you're already on the lookout for abuse for whatever reason, it would be hard to pick up on.” (Nurse) Q11: “As they're being asked the question again, the story keeps changing, so that's like a huge red flag and I think whoever the clinician is, whether it be a nurse, N.P., or the doctor, as they're viewing the patient's history and everything and chart and how many times have come to the E.R. with certain complaints so on and so forth, how the interaction is with the parent, how the interaction is with the child based on the developmental level; that's when we will start to like—antennas will go up, and we'll start having a conversation, nurse with doctor, doctor with nurse, doctor with doctor, and—nurse with nurse, about the situation.” (Nurse) Q12: “To identify child abuse is like look for kids who are frequently injured, or you get one x-ray and you see a healing injury that they didn't tell you about, or like they bounce around from hospital to hospital. It's like these like textbook indicators that you're not going to pick up on one E.R. visit necessarily, like coming to my E.R. doesn't like—I have no idea if they've gone to 6 other E.R.s this month.” (Nurse) Q13: “It definitely could be missed if it's their specific chart and this child is coming back for the same thing, it just depends how much the provider seeing them looks back in the note. Like oftentimes we'll look back at a recent admission and look at the summary and scroll back but it definitely could be missed and if the sibling for sure has come for a similar complaint that would be missed because there's no connection.” (Physician)

Abbreviation: EHR: electronic health records.

In addition, clinicians discussed how differences in daytime hours of admission and the volume of patients in the ED can hinder their ability to detect abuse and neglect adequately. For example, when ED patient volumes are high, clinicians have limited time with each patient and they must prioritize care between patients (Q3). Clinicians repeatedly recognized resource limitations, specifically that night shifts have fewer in-person staff (such as social workers and child abuse pediatricians), limiting options to consult with experts within the ED (Q4).

Next, clinicians highlighted several challenges of diagnosing child abuse and neglect due to ambiguity concerning the symptoms and/or lack of symptoms. In (Q5), a physician shares their concerns that physical exams cannot guarantee a valid detection of sexual abuse. For this reason, it is beneficial to understand social and family history and, when warranted based on screening questions and history, conduct tests and exams to determine the risk of abuse and neglect (Q6). Eight of the participants (40%) indicated the subtle nature of information, especially when it comes to neglect. Clinicians stated that neglect does not often have physical signs or symptoms that are easily detectable in the ED, and clinicians must rely on reports from parents and children who may provide conflicting accounts of the home environment (Q7 and Q8). Abuse is a sensitive and often secretive matter that can require time and a positive rapport with clinicians before parents disclose this information (Q9).

Finally, scattered and fragmented information in the EHR made it challenging for clinicians to create a comprehensive picture of abuse and neglect. EHR data can help clinicians understand patient history, including the number of previous ED/hospital visits, past complaints, and concerns regarding the parent–child relationship. Due to the large number of clinical notes and lack of access to outside hospital records, clinicians may have difficulty identifying patients at high risk for child abuse or neglect within the notes (Q10–Q12). Furthermore, the lack of medical history and social data on siblings or other family members limits clinicians’ ability to understand the families’ medical history (Q13).

Theme 2: different documentation styles in EHR

This theme highlights differences in clinical documentation by health discipline (physicians, nurses, social workers) that are important to consider when developing an AI-based clinical decision support system to identify potential child abuse and neglect by processing pediatric ED EHR data. These differences further articulate the challenges of diagnosis and care and the variety of vocabulary in clinical notes (Table  2).

Table 2.

Different documentation styles in EHR

Examples of language used in clinical notes describing child abuse and neglect Q14: “You can't write, ‘suspected child abuse’.” (Nurse) Q15: “Write bruise or lacerations, or healing wounds noted in generalized body area, or noted on arms and legs.” (Nurse) Q16: “I know they could put ‘positive for abuse’ or something to that extent in their note.” (Social work) Q17: “You know ‘concern for bruising’ or ‘abnormal bruising’ or ‘bruising’ and then I’ll describe, ‘patient is well-appearing with bruising on their anterior shins which can be typical of 4-year-old behavior. No other bruising anywhere else.’ I'll kind of describe my thinking process and then where that leads me like social work, please follow-up, labs that's pretty much it.” (Physician) Q18: “I might write, ‘evaluation for suspected physical abuse,’ or ‘suspected physical abuse,’ or ‘physical assault.’ That's if there are findings and somebody's obviously got findings, it's more physical assault. And the same is true of sexual evaluation.” (Physician) Q19: “Thing we say the most is NAT, non-accidental trauma for pediatrics, and any referrals to ACS (Administration for Children's Services) or CAC (Child Advocacy Center) is our child abuse team. And I don't know, maybe there might be some stuff documented in our social history section that would be relevant, but social work consults often.” (Physician)

Documentation style differences Q20: “Keeping the documentation concise and uniform if you're not the one collecting the full history, I don't think it's necessary to document a ton on that specific patient; it just can complicate that documentation.” (Nurse) Q21: “The abuse portion of it, I feel is unstructured. You kind of have to pull the information from different places to put that together” (Nurse) Q22: “Especially when it comes to abuse and neglect—you want to be as descriptive as possible. You want to document everything. You want to be very meticulous and then you also want to describe your concerns and the interventions.” (Physician) Q23: “The only thing that's really in the structured part is the exam (of child abuse and neglect).” (Physician) Q24: “In the E.R., our assessments typically look like one long narrative because it's one visit, so I open with their presenting factor or their chief complaint and then provide information that I was provided from the doctors as to why I’m being consulted on the case. I will then provide demographic information on the patient, um, and I typically follow a biopsychosocial.” (Social Worker)

Type or section of medical note Q25: “The pediatric (social work) assessment, and then there's a narrative that you put in, so part of the narrative is everything you talked about.” (Social work) Q26: “You'll see it (suspicion of child abuse and neglect) —usually they'll go under the provider note. The provider note will have all of it (details) and anybody—nurses, they can see the provider notes.” (Nurse) Q27: “Whatever concern about the mechanism of injury, and part of their plan of care would be reach out to social work or whatever the case may be, so it would probably be first documented in the provider's note and then in any follow-up like the social work note.” (Nurse) Q28: “So, there's like a ‘history of present illness’ part, so I would document let's say this is someone that, from getting the story, I find out there's concern that the child is being abused at home. I would write it in there um and then again you would see it again in my assessment and my plan.” (Physician)

Abbreviation: EHR: electronic health records.

All nurses interviewed stated that their language used for documentation and diagnosis is based on medical assessments and observations rather than other clinicians who may provide more contextual information. For example, nurses usually did not use the terms “abuse” and “neglect” in their clinical notes. They preferred using terminology that is more “objective” such as “bruises” or “healing wounds” with a description of the body area (Q14 and Q15). Conversely, social workers in the ED used vocabulary such as “abuse” and “neglect” in their clinical notes (Q16). Although physicians provided evaluations and routinely consulted with social workers, they were hesitant to incorporate the terms “abuse” and “neglect” in their clinical notes. For example, 3 physicians described how they used phrases such as “suspicion of abuse and neglect,” “concern for,” “evaluation for suspected physical abuse,” and “non-accidental trauma” (Q17–Q19).

In addition to the various descriptions of abuse and neglect in clinical notes, clinicians also described having different documentation styles and used both unstructured and structured clinical notes. For example, nurses described that they documented suspicion of abuse and neglect in clinical notes, in a brief and simple manner (Q20 and Q21). Alternatively, physicians used both clinical notes and structured EHR fields and tended to provide detailed reports regarding interventions and concerns in order to thoroughly document the suspicious cases (Q22 and Q23). Social workers tended to use long narratives in clinical notes for describing child abuse and neglect. Social workers explicitly documented their suspicion, reports from other disciplines, and the patients’ detailed social and psychological factors (Q24).

Finally, clinicians provided details on the specific sections and types of clinical notes where they usually documented concerns about child abuse and neglect. For example, a social worker described how social workers in the ED document their interactions with their clients in a “pediatric social work assessment” clinical note (Q25). Nurses highlighted that child abuse and neglect are first documented in the physician notes. Nurses also articulated that social work notes are sufficient for obtaining information regarding patients’ follow-up and plan of care (Q26 and Q27). One physician also described that they would document abuse both in the “history of present illness” and the “assessment and plan” (Q28).

Theme 3: identifying potential racial bias through documentation

Study participants suggested that although racial bias exists in healthcare, identifying it in clinical notes is challenging. We explored mechanisms through which implicit and explicit racial bias may be reflected in the EHR in relation to the evaluation and treatment of child abuse and neglect (Table  3). Clinicians indicated that racial bias is challenging to detect in the EHR because clinicians tend to be very cautious in how and when they use specific language when documenting concerns for child abuse and neglect (Q29 and Q30).

Table 3.

Identifying potential racial bias through documentation

Challenges identifying racial bias through EHR Q29: “I think it's just hard to quantify how those [racial bias] are documented, but I would say certainly like they're filtered through all the providers brains who have biases associated with it and I think the huge bias is that if it's not reported or the suspicion [child abuse and neglect] is not there, they're not going to write it even if they crossed their mind or they thought it. It's just not going into the chart.” (Physician) Q30: “That's tough I don't know because like I mean obviously like things that are documented in every patient are like demographics so their race will be documented but specifically in a physical exam or an assessment and plan I can't think of—I can't think of any specific terminology other than if I was assessing a chart and I saw this kid was you know non-white or you know or you know minority you know race and then read through the case—read through a case, seen if it's reported or not, and then you know looked at their race, I guess that's kind of how I would look at—be blind to the race while I read the chart and then make my determination about report or not and then look at what the race was. That's the only way I can really think of how I would because I don't think—I’m trying to think like how I document—I don't know if there would be any—anything else I could look at.” (Physician)

Potential clinician bias in quality of care Q31: “I assume people document differently, I mean if I'm right that people are treating them differently then I assume they document differently, because I think the documentation has to justify the treatment.” (Physician) Q32: “I definitely think there is 100 percent disparity with injuries and a higher level of reporting (to child services) of African-American, Hispanic versus Caucasian 100 percent.” (Nurse) Q33: “So, I think that there is a kind of stereotype as to what a family who has a child who's being abused looks like. Right? And it's basically kind of what is portrayed in society and the media and the news right? (…..). So, when a specific demographic, let's say. may that be like a person of color, Black or Hispanic, young mom with several children come in your index of suspicion for non-accidental trauma might be higher than a Caucasian family mom and dad who come in who is a lawyer or you know what I mean? So, there's definitely bias that goes into it.” (Physician) Q34: “I would say that they might be treated differently or there might be a more hyper-vigilant eye on them if there is a concerning factor, so if this is a young mom, 19 18 years old who presents with their infant who's fallen, she doesn't have a lot of social support or whatever, or a baby that comes in with failure to thrive, then their age is I wouldn't say that they're treated differently unfairly though I would say that their age is a factor to consider and definitely plays a role in whatever intervention we would put into place.” (Social Work)

Q35: “Where I think someone was treated unfairly. (….) I would say that for example, oh, the language one is a big thing, right? Like if you are having a hard time communicating you might not spend as much time trying to communicate to them as someone who speaks your own native language.” (Physician)

Specific language used to detect potential racial bias in EHR Q36: “I mean of course as soon as you start to go in a certain direction you start using language like ‘the parent states,’ ‘the parent alleges,’ that you know describing what the parent says rather than describing what happens right so the normal way that we describe things if a kid falls going down the slide and breaks his leg we say, ‘the child was going down the slide and fell and broke his leg,’ whereas if we doubt, we say that parent states it happened, so I mean that would be the first thing.” (Physician) Q37: “If the chart mentions race, that's probably bias. It's 90 plus percent of time irrelevant for emergency care, especially in pediatrics, maybe more, so maybe 95% if the chart mentions race. There are certain words that I think are associated with, you know, bias. It's not specifically racial, but I think when we say, ‘noncompliance’ like we talked about patients being ‘non-compliant’. When we talk about ‘pain seeking, drug seeking behavior’ those kinds of words, they tend to be about specific populations.” (Physician) Q38: “I've seen notes that mentioned ‘the difficult patient’. To me personally I find that to be inappropriate. (….) the use of the term ‘difficult’ in a chart.” (Physician)

Abbreviation: EHR: electronic health records.

Despite the challenges mentioned above, several clinicians (55%) provided important insights into how clinical notes can reflect implicit and explicit racial biases. For example, one physician mentioned that it is possible if documentation justifies the treatment, the EHR may reflect differences in care (Q31). In (Q32), a nurse described how she believes that Black and Latinx families have higher reporting of abuse to child services than white families. Clinicians further described how abuse and neglect are stereotyped and portrayed in society and media, which may impact real-world reporting. For example, 2 interviewees, a physician and a social worker, describe that families of color or young parents may be more often suspected of abuse (Q33 and Q34). Other quality-of-care gaps may emerge due to patients and clinicians speaking different languages, thus leading to communication difficulties and resulting in shorter assessments (Q35).

Clinicians stated that there are examples of potential racial bias reflected by vocabulary used within clinical notes. Clinicians describe how the language used in clinical notes to describe potential abuse could determine if they believe or doubt the patients’ and families’ accounts (Q36). Furthermore, the mention of the patient’s race in clinical notes is usually irrelevant for emergency care and may reflect racial biases. In addition, clinicians witnessed notes that mention words such as “noncompliance,” “pain seeking,” “drug-seeking behavior,” and the “difficult patient.” Physicians suggest that this language tends to be about people of color and it is inappropriate for medical evaluation and documentation (Q37 and Q38).

DISCUSSION

This qualitative study identifies several barriers to detecting child abuse and neglect in ED setting and highlights several important considerations for generating a phenotype for child abuse and neglect using EHR data. The findings in this study emphasize the importance continuity of care have on the diagnosis of child abuse and neglect. In EDs clinicians have limited time for each patient and they often do not see their patients more than once. This lack of continuity of care impacts clinicians’ ability to assess and diagnose different patterns of abuse and neglect. Other studies support these findings and suggest that clinicians tend not to detect child abuse and neglect cases during a first ED visit, leading to a delayed or missed diagnosis of child abuse and neglect,¹⁹ potentially resulting in increased morbidity and mortality.²⁰

In this current study, clinicians described the challenges of diagnosing child abuse and neglect during periods of high patient volumes in the ED, due to time limitations. Surprisingly, one study found that hospitals with low admission volumes were less likely to assess child abuse and neglect with screening tools such as skeletal surveys and were less likely to admit children with suspected abuse.²¹ These findings require further exploration with larger quantitative data.

Clinicians also described the challenges in evaluating abuse and neglect during night and weekend shifts, due to limited availability of clinical resources such as lower number of clinicians. Moreover, due to a lack of staff, there are fewer possibilities for consultations and teamwork to assess potential abuse and neglect. These findings are supported by previous research that underlines the barriers to detecting child abuse and neglect due to time constraints, lack of familiarity with the patient, and built-in working challenges in EDs,²² such as lack of interdisciplinary work efforts.²³

Findings from this study also highlights difficulties in diagnosing abuse and neglect due to subtle information and ambiguity concerning the symptoms and signs of potential child abuse and neglect. Clinicians state that diagnosing abuse and neglect requires a deep understanding of social and family history combined with a thorough physical examination of the patient. Neglect is a well-disguised phenomenon, and identifying neglect requires more time spent with patients and their families. Recent research identifies clinicians’ desire to believe the patient or family, thus resulting in a failure to recognize signs of abuse and neglect. To resolve some of these issues, other studies suggested including an educational approach for improving identification by reviewing reported cases and following up with child protective services.²²

Our findings indicate that social, family, and medical histories are scattered and fragmented within the EHR. Clinicians state that there is no possibility to easily retrieve information in the EHR regarding siblings or other family members, thus resulting in a lack of context for further understanding potential abuse and neglect. The significant number of clinical notes impact abuse evaluation as well, as EHR data usually are not presented and clearly organized and, therefore, are hard to navigate and understand. Poor usability of EHRs can result in difficulties for clinicians in locating essential information for detecting and diagnosing patients suffering from health and social problems,²⁴ such as child abuse and neglect. To improve the usability of EHRs and the efficiency of child abuse and neglect assessments, we recommend the involvement of multiple stakeholders to redesign and improve workflow surrounding the collection and storage of medical data in the ED.²⁵

Clinicians in this study confirm past research^26,27 and describe how the lack of interoperability of EHR data affects their ability to develop a comprehensive understanding of abuse and neglect. There are notable gaps in information transfer and remote access to medical data in the United States, preventing clinicians from obtaining and exchanging medical information between different hospitals and EDs.²⁸

Our study identified that various health professionals in one ED document child abuse and neglect differently. For example, nurses are more likely to consult with other clinicians and, therefore, will use strict and “objective” vocabulary in their clinical notes. On the other hand, social workers provide richer context and in-depth description of social and family history and plan of care in their clinical notes. Physicians tend to provide detailed information regarding injuries related to potential abuse and neglect. These interdisciplinary differences in documentation are expected as they reflect different roles within the clinical environment.^29,30

Findings in this study suggest that examining EHR data can expose racial disparities in healthcare. Other recent studies also showed that EHR data can reveal racial bias by demonstrating how clinicians document. For example, one recent study found linguistic bias in the records of black patients; clinical notes written about Black patients had higher odds of containing “judgment words” suggesting doubt (eg, “claims” or “insists”) and linguistic features suggesting disbelief (eg, had a “reaction” to the medication), compared to notes of white patients.³¹

However, most EHR-date based risk systems are not sensitive to racial disparities. For example, in a recent natural language processing study to identify child abuse within the EHR, no efforts were implemented to limit the potential impact of racially biased language on identifying patients at risk.⁹

Finally, our research indicates that EHR data can potentially expose inequalities in the ED experienced by Black and Latinx patients; clinicians reported that these communities might be described using different language compared to white patients. These inequalities do not explicitly concern child abuse and neglect, but they can illustrate how racial bias is a source of pediatric health care disparities,³² potentially impacting the evaluation of abuse and neglect.

Implications for EHR-based phenotype development

Our results have several implications for developing an EHR-based phenotype of suspected abuse and neglect. First, more efforts are needed to generate continuous understanding of patients’ healthcare use, including previous ED and hospital visits and ideally outpatient data. Second, abuse and neglect risk phenotype might be affected by timing of documentation and fluctuating patient numbers, since clinicians’ documentation length might decrease with increased workload. Thirdly, phenotyping efforts that use EHR data will need to incorporate documentation produced by multidisciplinary clinicians while being sensitive to interdisciplinary differences in the way clinicians document. Fourth, particular attention needs to be paid to efforts that can help reduce overconsideration of child abuse and neglect among communities of color. For example, such efforts can use natural language processing to help identify the language clinicians tend to use when describing certain patient populations. The effect of such language can be adjusted when developing risk phenotypes.

These recommendations must be implemented in conjunction with the facilitation of a better understanding on the part of medical staff of the social and cultural nuances of the populations served. Hospitals and clinicians must create initiatives—such as cultural competency training, diverse hiring practices, and the deliberate elevation of the voices and experiences of marginalized clinicians and patients in healthcare—in order to best recognize and handle these nuances.

Limitations

This study has several notable limitations. First, this was a qualitative study conducted in one ED of a large urban hospital and our results might not be generalizable to other settings or locations. In future research, the involvement of additional EDs and clinicians is essential. The number of involved interdisciplinary interviewees was also low for some disciplines—this might affect results’ generalizability. Second, the study raised several sensitive questions (eg, racial bias) that can be challenging to address only via interviews, and further studies exploring these issues with additional methodologies (qualitative and quantitative) are warranted. Third, this study did not discuss differences in work experiences among clinicians and how they impact diagnosis and documentation of child abuse and neglect within the ED. Future research is required to further understand these differences.

Implications for further research

In general, further research is needed in 3 areas: (1) generating a comprehensive understanding about how to process and synthesize the various sources of information related to child abuse and neglect in the EHR; (2) addressing issues associated with the ED setting as a source of emergent rather than continuing care; and (3) exploring how racial bias is being captured in EHR data.

CONCLUSIONS

This was the first qualitative study to gain insights about generating an EHR-based phenotype that will help to identify children at risk for abuse and neglect in the ED setting. Our findings highlight several challenges of building a risk phenotype, including lack of data about previous service use, interdisciplinary differences in documentation, and potential for racially biased language that might be used by clinicians. Further research is needed to validate these findings and integrate them into creation of an EHR-based risk phenotype.

FUNDING

This project was supported by the Seeds Grant Fund at the Data Science Institute at Columbia University for innovative and interdisciplinary research (to AL, MT, and DP). In addition, AL was funded by the Haruv Institution in Israel for researching child abuse and neglect.

AUTHOR CONTRIBUTIONS

AL wrote the draft and facilitated group discussion. AB, KC, NA, SS, DP, and MT reviewed and refined the draft and finalized the article.

CONFLICT OF INTEREST

The authors do not report any potential conflicts of interest. Each author has confirmed compliance with the journal’s requirements for authorship.

DATA AVAILABILITY

The data underlying this article cannot be shared publicly to maintain the privacy of individuals that participated in the study. The data will be shared upon reasonable request to the corresponding author.