Estimating the Prevalence of Intimate Partner Violence at an Urban Hospital Before and During the COVID-19 Pandemic Using a Novel Natural Language Processing Algorithm

Abstract

The impact of COVID-19 on intimate partner violence (IPV) in the United States is still relatively unknown, although some early data demonstrate that cases of IPV increased during COVID-19. The objective of this study was to measure the prevalence of IPV before and during the COVID-19 pandemic in a southeastern urban hospital. We performed a retrospective analysis of IPV encounters at a single high-volume Level I trauma hospital. IPV encounters were identified through a novel natural language processing algorithm using IPV-related words and phrases within unstructured clinical notes. IPV encounters from February to August 2019 (pre-COVID-19 period) were compared with encounters from February to August 2020 (COVID-19 period). The IPV visit rate during the COVID-19 period was higher than that during the pre-COVID-19 period (0.82% of all visits in 2020 vs. 0.72% of all visits in 2019). The number of IPV encounters for patients with no prior IPV visits was higher in 2020, whereas the number of revisits, patients with prior IPV encounters, was lower in 2020. There was an increased incidence of IPV during the COVID-19 pandemic with an increase in the number of patients presenting with first time IPV encounters. Future hospital and community pandemic preparedness protocols must include expansion of screening, resource allocation, and protective policies for those in unsafe situations.

Introduction

Early in the SARS-CoV-2 (COVID-19) pandemic, public health experts postulated that pandemic conditions would exacerbate intimate partner violence (IPV) (Viero et al 2021; Evans 2020). Increasingly the evidence suggests that to be true, with an observed increase in calls to domestic violence hotlines and domestic violence-related crimes in the United States (Evans et al 2021; Piquero et al 2021). Yet, the extent to which the pandemic has impacted health-seeking behaviors and the incidence of IPV among those seeking medical care in U.S. health systems is still relatively unknown.

Recent studies have suggested an increase in the incidence of IPV among individuals seeking medical care during the pandemic, with a higher severity of injury due to IPV (Gosangi et al 2021). Because individuals experiencing IPV frequently seek medical care as a result of injuries, it is important to understand how the pandemic impacted IPV and related health-seeking behaviors.

The measurement of IPV in health care settings is a significant challenge worsened by the pandemic. Before the pandemic, the documentation and coding of IPV were inadequate leading to poor identification in electronic health records (EHRs). In fact, one study identified that IPV-specific International Classification of Diseases (ICD)-10 codes identified only one-fourth of all female patients who presented to the emergency department (ED) for IPV-related care (Schafer et al 2008). Thus, identifying IPV by ICD-10 codes exclusively has been proven to result in inadequate measurement.

During the pandemic, the ability to recognize and accurately measure the number of individuals seeking care was made more challenging since hospitals were operating at or over capacity and with limited staffing (Butler et al 2020; Cavallo et al 2020). As a result, routine IPV screening practices were often suspended during COVID-19 due to the increased demand on health systems and health care providers (Wu, personal communication, July 1, 2020).

IPV continues to be a serious and preventable public health problem, affecting >10 million people annually in the United States (CDC 2021). The need for accurate surveillance methods is critical, especially in health care settings where interventions can be implemented at the point of care. For example, the Cardiff violence prevention model has been successfully implemented in health system EHRs to screen for violence-related injury (CDC 2018). Implementation of the screening tool has assisted with targeted violence prevention interventions in areas with high crime rates (Mercer et al 2020).

Although screening tools such as the Cardiff model are not yet universal, additional informatics-based strategies can aid in the identification of IPV. The application of natural language processing (NLP) has been successfully utilized to identify specific medical conditions, including difficult to identify, socially stigmatized, and/or uncommon conditions (Blosnich et al 2018; Juhn and Liu 2020). NLP is a technique that allows for extraction and analysis of large amounts of clinical data including words and phrases from unstructured or free text data contained in EHRs (Juhn and Liu 2020; Koleck et al 2019).

This technique has been utilized to identify sexual trauma among individuals seeking care within the Veterans Health Administration and to develop a screening algorithm to identify and predict child abuse (Annapragada et al 2021; Divita et al 2017). A recent study describes the use of a novel NLP algorithm to identify IPV using unstructured EHR notes (Tabaie et al 2022).

The main purpose of this study was to apply a novel NLP algorithm to measure IPV prevalence before and during the COVID-19 pandemic in an urban hospital setting. Secondarily, we explored IPV-related encounters before and during COVID-19 to determine whether visits during the pandemic were first time IPV encounters (patients with no prior IPV encounters) or revisits (patients with a history of IPV-related encounters). Finally, we sought to understand any potential differences in demographic-related factors (e.g., age, gender) or hospital-related factors (e.g., disposition, mortality, chief complaints) among those experiencing IPV.

Methods

Study design and setting

This study utilized a retrospective cross-sectional analysis of all ED encounters at a large southeastern U.S. urban academic Level I trauma and emergency care center with >140,000 annual ED visits. The hospital serves as a public safety-net hospital for two of the largest counties in Georgia, serving a largely underinsured population. This study was approved by the Emory University institutional review board.

Study sample identifying true positive IPV encounters

Current methods of identifying IPV cases using EHR data are primarily limited to using ICD-10 codes. Using this method, true IPV encounters are severely undercoded (Schafer et al 2008); thus, relying on ICD codes exclusively misses a large number of encounters. As a result, we employed the use of NLP to accurately identify true IPV encounters from 2012 to 2020.

All ED patient encounters from January 2012 to August 2020 were included in the data set. Starting in January 2012, all ED encounters transitioned from paper charting to EHR using Epic. This time period was selected to maximize the amount of data with which to develop the algorithm. The data obtained from each ED encounter included unstructured data from ED notes including all provider, nursing, and social work notes as well as chief complaint, ICD-9 or ICD-10 code, disposition, length of stay (LOS), and status (Tabaie et al 2022). These data were used to develop an IPV algorithm that accurately identified positive IPV cases from 2012 to 2020.

The NLP algorithm was developed using specific IPV-related words and phrases contained within unstructured clinical notes. The IPV-related words were predetermined by the study team using IPV terms derived from existing literature and from clinician expertise (e.g., mechanism-related terms, assault by partner). After application of the algorithm, positive IPV cases were manually reviewed by members of the study team (AJZ) to further refine the algorithm components and to determine accuracy of the encounters labeled as positive IPV encounters.

The initial algorithm included autopopulated IPV screening questions (regardless of whether or not screening was completed), encounters in which patients were noted to have a history of IPV but the current encounter was not for an IPV-related issue, and encounters in which the patient denied a history of IPV. Modifications were made to the algorithm to omit these false positives (Chapman et al 2001). This iterative process of manually reviewing encounters identified by the algorithm as positive cases to assess for accuracy was continued until no false positive cases were identified (Tabaie et al 2022). In total, manual review was conducted on 25% of identified IPV cases.

Comparison of IPV prevalence before and during the COVID-19 pandemic

To explore the impact of COVID-19 on IPV, we reviewed positive IPV encounters identified by the NLP algorithm from February–August 2019 (pre-COVID-19 period) to February–August 2020 (COVID-19 period). We evaluated for differences in the number of IPV visits, encounters by patients with no prior IPV encounters, encounters by patients with prior IPV encounters, age, gender, patient status (i.e., mortality), disposition status, LOS, chief complaint, and admission diagnosis.

Statistical analysis

Statistical analysis was conducted to compare differences between the pre-COVID-19 (2019) and COVID-19 (2020) study periods. For the numerical features such as age and LOS, a Wilcoxon test was performed. For mortality, a Fisher's exact test was performed; for patient disposition a chi-squared proportional test was applied. For patient disposition, only “discharged” and “admitted” statuses were considered as there were a negligible number of encounters in which patients eloped or left against medical advice.

Results

During the 2012–2020 study period, there were 1,064,735 ED encounters included in the analysis representing 405,303 patients. Unstructured notes from these encounters were used to develop the IPV algorithm, defined as encounters in which a predetermined IPV-related term was identified. During this 9-year time period, a total of 7399 positive IPV encounters were identified using the algorithm representing 5975 patients.

Patient demographics before and during the COVID-19 pandemic

When comparing the prepandemic (2019) and pandemic (2020) study periods, the median age of patients experiencing IPV was 33 years for both study periods (Table 1). Most patients experiencing IPV were female in both 2019 and 2020 (n = 503, 85.1% 2019, n = 410, 82.5% 2020). The disposition of patients was relatively unchanged during 2019 and 2020 with the majority of patients being discharged (n = 407, 68.9% 2019, n = 363, 73% 2020). There was no statistically significant difference in mortality (n = 7, 1.2% 2019, n = 3, 0.6% 2020) or LOS (n = 6 h 2019, n = 5.4 h 2020) between the two time periods.

Table 1.

Patient Demographic and Clinical Characteristics Before and During the COVID-19 Pandemic

Characteristics	February–August 2019	February–August 2020	p
Age, median (IQR), year	33 (25–42)	33 (26–43)	0.19
Gender, n (%)
Male	88 (14.9)	87 (17.5)	0.24
Female	503 (85.1)	410 (82.5)	0.24
Disposition, n (%)
Admission	184 (31.1)	134 (27)	0.13
Discharge	407 (68.9)	363 (73)	0.13
Mortality, n (%)	7 (1.2)	3 (0.6)	0.36
LOS, median (IQR), hours	6 (3.5–9.4)	5.4 (3.4–9.5)	0.29
Total number of IPV cases, n (% all visits)	591 (0.72)	497 (0.82)	0.02

IPV, intimate partner violence; IQR, interquartile range; LOS, length of stay.

IPV incidence, recurrence, and chief complaints before and during the COVID-19 pandemic

The overall IPV visit proportion was higher during the COVID-19 pandemic period than during the prepandemic period (Table 2). In 2019, IPV visits represented 0.72% (n = 591, total n = 82,650) of all visits compared with 0.82% (n = 497, total n = 60,670) (p = 0.02) of all visits in 2020. Notably, the number of IPV encounters for patients with no prior visits for IPV was significantly higher during the pandemic period in 2020 (n = 361, 0.59%) versus 2019 (n = 203, 0.25%). The number of revisits, defined as individuals with a prior presentation for IPV presenting with IPV, was lower in 2020 (n = 72, 0.12%) than in 2019 before the pandemic (n = 314, 0.38%).

Table 2.

Intimate Partner Violence-Related Encounters in a Hospital Setting Before and During the COVID-19 Pandemic

IPV-related encounters	February–August 2019, n (%)	February–August 2020, n (%)	p
Prevalence	591 (0.72)	497 (0.82)	0.02
New visits (IPV MRNs with no prior IPV visits)	203 (0.25)	361 (0.59)	<0.001
Revisits (IPV MRNs with prior IPV visits)	314 (0.38)	72 (0.12)	<0.001

MRN, medical record number.

Before the pandemic in 2019, the most frequent chief complaints among patients experiencing IPV were “trauma,” “psychiatric evaluation,” and “alleged domestic violence”; the most frequent chief complaints during the pandemic in 2020 were “trauma,” “alleged domestic violence,” and “sickle cell pain crisis” (Fig. 1). The most common admission diagnoses for both periods were similar, “unspecified chest pain” and “unspecified abdominal pain.”

FIG. 1.

Chief complaints from IPV encounters pre-COVID-19 and during COVID-19. IPV, intimate partner violence.

Limitations

As our algorithm relied on identification of specific IPV-related words and terminology in unstructured notes, it is possible that we did not identify cases in which patients chose not to disclose IPV as the reason for the encounter. Presumably, this limitation would be consistently present throughout the study period and would theoretically not impact comparisons from 2019 to 2020. In addition, the algorithm identified IPV-related words that may have been used in notes outside of the context of IPV. For example, “domestic dispute” may have identified a case in which an altercation occurred between individuals that may or may not have been related to IPV (unless further clarified in the note).

Again, we would expect this limitation to be relevant to the entire data set and thus would not impact comparisons across time periods. When evaluating for new IPV encounters and revisits, only prior encounters within the same health system were available. It is possible that individuals presented to other health systems for IPV-related visits, and these records were not available to us resulting in an underestimate of repeat IPV. We were unable to disaggregate data based on race, sexual orientation, gender identity, or insurance status given the limitations of data available in the EHR.

This limitation is important given that many of these are risk factors for IPV and also represent structural vulnerabilities that may be exacerbated by COVID-19 (Capaldi et al 2012; Lo et al 2021). Finally, results are applicable only to the health system included in the study, which limits generalizability to other health systems.

Discussion

Using a novel NLP algorithm, our findings demonstrate an increase in the proportion of IPV-related ED encounters during the COVID-19 pandemic. We also identified an increase in the number of IPV encounters by individuals with no prior IPV visits and a decrease in IPV-related encounters by individuals with prior IPV-related encounters. These findings suggest that individuals with no history of IPV were impacted negatively by COVID-19 and pandemic-related movements and/or stressors, resulting in escalating interpersonal conflict and first time IPV-related injury. The decrease in revisits is surprising as a history of IPV is a risk factor for future IPV.

It is possible that those with a history of IPV-related encounters sought care in a health setting outside of the health system included in this study, or were potentially unable to seek care during this time period. The reasons for this are likely multifactorial but could be related to COVID-19 movement restrictions, fear of COVID-19 infection, and/or partner-initiated restrictions as individuals may have been forced to quarantine with abusive partners with limited access to support systems. Interestingly, we did not observe a statistically significant change in demographic or hospital-related factors during the COVID-19 pandemic despite the many strains on hospital systems during this time.

The findings in our study are similar to that of Ebert and Stenert demonstrating an increase in IPV during a similar COVID-19–specific time period (April–May 2020) in Germany (Ebert and Steinert 2021). The study methods differed as their study was an online survey study based on self-reported data, however, they also found an increase in IPV during COVID-19. A study conducted in a U.S. hospital system evaluating IPV-related injury from radiological findings demonstrated a decrease in IPV-related encounters during the COVID-19 pandemic but an increase in the incidence of physical violence among IPV encounters (Gosangi et al 2021). Notably, the IPV-related encounters were determined from referral data to an IPV prevention program, and the number of encounters was only compared with referrals from prior years, not overall hospital visits. As many hospital systems saw a decrease in total visits early during the COVID-19 pandemic, it is difficult to interpret the overall proportion of IPV-related encounters without comparing with total visits. Finally, several studies have demonstrated that IPV increases during emergencies, including pandemics, natural disasters, and humanitarian emergencies (Bell and Folkerth 2016). After the Ebola virus disease (EVD) pandemic in Sierra Leone, cases of all forms of gender-based violence increased compared with before the EVD pandemic, including reported cases of domestic violence (UNDP 2014). Relatedly, the prevalence of IPV increased after Hurricane Katrina especially among individuals who were directly affected by hurricane-related stressors (i.e., displacement, unstable housing, food shortages) (Anastario and Shehab 2009; Schumacher et al 2010).

The ongoing COVID-19 pandemic has and will likely continue to exacerbate IPV. Those experiencing IPV are forced to balance the risk of remaining in an unsafe relationship with the risk of seeking help and potential exposure to COVID-19. Given these issues, it is possible that only individuals who were seriously injured sought care. Furthermore, the pandemic resulted in significant economic hardship including but not limited to food, housing, and employment instability—all conditions that are known to worsen interpersonal conflict (CBPP 2021). Simultaneously, individuals suffered from limited access to much needed support systems, social services, and access to medical care during the pandemic (Wood et al 2020). For those who did access care, health care providers reported significant challenges in care provision (limited time/excessive workload, decreased utilization of screening tools) and multiple barriers to safe discharge planning for those experiencing IPV during the pandemic (Hendrix et al 2022). It is critical that future pandemic preparedness protocols are considerate of mandates that may inadvertently harm individuals, recognizing that quarantine and social isolation may intensify unsafe conditions. Concurrently, protocols should prepare for expansion of resources that can be quickly and easily accessed for those living in unsafe conditions. This is especially true for health systems as hospital-based interventions can be implemented at the time of care.

The need for universally implemented IPV screening tools in health systems can serve a critical role in the identification and subsequent support for those experiencing IPV (ASPE 2021). Hospital administrators should ensure that screening-related practices continue during emergencies since such conditions are a trigger for IPV-related injury. It is also critical that hospitals prepare additional IPV-related resources, working in collaboration with community organizations to develop safety planning and safe disposition that accounts for pandemic-related restrictions. As individuals may have significant barriers in accessing much needed care for IPV-related injuries, it is important to explore expanded opportunities for outreach for those who are not able to leave their homes. This could include the thoughtful use of telemedicine and mobile outreach to access individuals living in unsafe conditions. Most importantly, it is essential that we understand the experiences of both survivors and the health care providers supporting them to guarantee that current and future interventions address their needs (Hendrix et al 2022).

Conclusion

This study highlights the need for robust support of a public health approach when addressing IPV during future pandemics. It is clear that pandemics and other health emergencies, including COVID-19, are a trigger for exacerbation of IPV. Future efforts must include appropriate surveillance mechanisms, increased funding for community resources, and implementation of policies that are protective or at least minimize harm for those living in unsafe situations.

Footnotes

Author Disclosure Statement

No competing financial interests exist.

Funding Information

This project was funded by the Emory University Woodruff Health Sciences Synergy Grant.

References

Anastario

, Shehab

. Increased gender-based violence among women internally displaced in Mississippi 2 years post-Hurricane Katrina. Disaster Med Public Health Prep, 2009; 3(1):18–26; doi: 10.1097/DMP.0b013e3181979c32.

Annapragada

, Donaruma

MMK

, Annapragada

, et al. A natural language processing and deep learning approach to identify child abuse from pediatric electronic medical records. PLoS One, 2021; 16(2):e0247404; doi: 10.1371/journal.pone.0247404.

Bell

, Folkerth

. Women's mental health and intimate partner violence following natural disaster: A scoping review. Prehosp Disaster Med, 2016; 31(6):648–657; doi: 10.1017/S1049023X16000911.

Blosnich

, Cashy

, Gordon

, et al. Using clinician text notes in electronic medical record data to validate transgender-related diagnosis codes. J Am Med Inform Assoc, 2018; 25(7):905–908; doi: 10.1093/jamia/ocy022.

Butler

, Wong

SPY

, Wightman

, et al. US clinicians' experiences and perspectives on resource limitation and patient care during the COVID-19 pandemic. JAMA Netw Open, 2020; 3(11):1–15; doi: 10.1001/jamanetworkopen.2020.27315.

Capaldi

, Knoble

, Shortt

, et al. A systematic review of risk factors for intimate partner violence. Partner Abuse, 2012; 3(2):231–280; doi: 10.1891/1946-6560.3.2.231.

Cavallo

, Donoho

, Forman

. Hospital capacity and operations in the coronavirus disease 2019 (COVID-19) pandemic—Planning for the Nth patient. JAMA Health Forum, 2020; 1(3):e200345; doi: 10.1001/jamahealthforum.2020.0345.

Center on Budget and Policy Priorities (CBPP). Tracking the COVID-19 Economy's Effects on Food, Housing, and Employment Hardships; 2021. Available from: https://www.cbpp.org/research/poverty-and-inequality/tracking-the-covid-19-economys-effects-on-food-housing-and [Last accessed: February 1, 2022].

CDC. Fast Facts: Preventing Intimate Partner Violence; 2021. Available from https://www.cdc.gov/violenceprevention/intimatepartnerviolence/fastfact.html [Last accessed: February 1, 2022].

10.

CDC. The Cardiff Model Toolkit; 2018. Available from: https://www.cdc.gov/violenceprevention/about/fundedprograms/cardiffmodel/toolkit.html [Last accessed: February 1, 2022].

11.

Chapman

, Bridewell

, Hanbury

, et al. A simple algorithm for identifying negated findings and diseases in discharge summaries. J Biomed Inform , 2001; 34(5):301–310; doi: 10.1006/jbin.2001.1029.

12.

Divita

, Brignone

, Carter

, et al. Extracting sexual trauma mentions from electronic medical notes using natural language processing. Prec Healthcare Through Inform, 2017; 245:351–355; doi: 10.3233/978-1-61499-830-3-351.

13.

Ebert

, Steinert

. Prevalence and risk factors of violence against women and children. Bull World Health Org, 2021; 99(6):429–438; doi: 10.2471/BLT.20.270983.

14.

Evans

DP.

COVID-19 and violence: a research call to action. BMC Womens Health, 2020; 20(1):249; doi: 10.1186/s12905-020-01115-1.

15.

Evans

, Hawk

, Ripkey

. Domestic violence in Atlanta, Georgia before and during COVID-19. Violence Gend, 2021; 8(3):140–147; doi: 10.1089/vio.2020.0061.

16.

Gosangi

, Park

, Thomas

, et al. Exacerbation of physical intimate partner violence during the COVID-19 pandemic. Radiology, 2021; 298(1):E38–E45; doi: 10.1148/radiol.2020202866.

17.

Hendrix

, Narasimhan

, Ripkey

, et al. Boiling water but there's no pop-off valve. Violence Against Women, 2022 (accepted).

18.

Juhn

, Liu

Artificial intelligence approaches using natural language processing to advance EHR-based clinical research. J Allergy Clin Immunol, 2020; 145(2):463–469; doi: 10.1016/j.jaci.2019.12.897.

19.

Koleck

, Dreisbach

, Bourne

, et al. Natural language processing of symptoms documented in free-text narratives of electronic health records: A systematic review. J Am Med Inform Assoc, 2019; 26(4):364–379; doi: 10.1093/jamia/ocy173.

20.

, Nguyen

, Drew

, et al. Race, ethnicity, community-level socioeconomic factors, and risk of COVID-19 in the United States and the United Kingdom. EClinicalMedicine, 2021; 38:101029; doi: 10.1016/j.eclinm.2021.101029.

21.

Mercer Kollar

, Sumner

, Bartholow

, et al. Building capacity for injury prevention: A process evaluation of a replication of the Cardiff Violence Prevention Programme in the Southeastern USA. Injury Prev, 2020; 26(3):221–228; doi: 10.1136/injuryprev-2018-043127.

22.

Piquero

, Jennings

, Jemison

, et al. Domestic violence during the COVID-19 pandemic—Evidence from a systematic review and meta-analysis. J Crim Just, 2021; 74:101806; doi: 10.1016/j.jcrimjus.2021.101806.

23.

Schafer

, Drach

, Hedberg

, et al. Using diagnostic codes to screen for intimate partner violence in Oregon emergency departments and hospitals. Public Health Rep, 2008; 123(5):628–635; doi: 10.1177/003335490812300513.

24.

Schumacher

, Coffey

, Norris

, et al. Intimate partner violence and Hurricane Katrina: Predictors and associated mental health outcomes. Viol Vict, 2010; 25(5):588–603; doi: 10.1891/0886-6708.25.5.588.

25.

Tabaie

, Zeidan

, Evans

, et al. A novel technique to identify intimate partner violence in a hospital setting. West J Emerg Med, 2022 (accepted).

26.

United Nations Development Programme (UNDP). Assessing Sexual and Gender-Based Violence during the Ebola Crisis in Sierra Leone; 2014. Available from: https://www.sl.undp.org/content/sierraleone/en/home/library/crisis_prevention_and_recovery/assessing-sexual-and-gender-based-violence-during-the-ebola-cris.html [Last accessed: February 1, 2022].

27.

United States Department of Health and Human Services, Office of the Assistant Secretary for Planning and Evaluation (ASPE). Screening for Domestic Violence in Health Care Settings; 2021. Available from: https://aspe.hhs.gov/reports/screening-domestic-violence-health-care-settings-0 [Last accessed: February 1, 2022].

28.

Viero

, Barbara

, Montisci

, et al. Violence against women in the Covid-19 pandemic: A review of the literature and a call for shared strategies to tackle health and social emergencies. Forensic Sci Int, 2021; 319:110650; doi: 10.1016/j.forsciint.2020.110650.

29.

Wood

, Voth Schrag

, Baumler

, et al. On the front lines of the COVID-19 pandemic: Occupational experiences of the intimate partner violence and sexual assault workforce. J Interpers Viol, 2020; 37(11–12): NP9345-NP9366; doi: 10.1177/0886260520983304.