Identifying Latent Patterns of Intimate Partner Violence Using Electronic Health Records and Association Rule Mining

Abstract

Intimate partner violence (IPV) remains a significant public health concern, yet its identification in emergency department (ED) settings is often challenging, as IPV-related ICD-10 are infrequently and inconsistently used. This study employs electronic health record (EHR) data and association rule mining to identify latent patterns indicative of IPV among female patients in ED. We applied the Apriori algorithm to identify associations between co-occurring health conditions in confirmed IPV survivors and non-IPV patients. The analysis revealed distinct patterns in both groups, with IPV survivors showing stronger links between physical injuries, mental health conditions (e.g., depression, suicidal ideation), and socioeconomic stressors. Non-IPV cases were primarily associated with anxiety disorders and family-related stressors. These findings highlight the potential of informatics-driven approaches to improve IPV detection by revealing subtle clinical signals that may notbe overtly disclosed. Our work paves the way for developing data-driven tools to enhance IPV screening and clinical decision-making in ED.

1. Introduction and Background

Intimate partner violence (IPV) represents a serious and persistent public health concern, with significant implications for healthcare delivery and clinical outcomes. IPV includes physical and sexual violence, threats, economic, and emotional/psychological abuse, or other abusive behavior as part of a systematic pattern of control and power perpetrated by one intimate partner against another.¹ IPV affects women disproportionately.² In the United States, one in four women experience IPV during their lifetime, and one in five require medical care due to IPV-related injuries or health consequences.³ Beyond acute injuries, IPV is associated with chronic conditions, including anxiety, depression, and chronic pain, driving increased healthcare utilization over time.^4-6

Despite its prevalence and clinical impact, IPV remains difficult to identify in emergency departments (EDs), where many IPV victims seek care. Victims frequently do not disclose IPV, due to stigma, fear of abusive partner, dependence on the perpetrator, or psychological trauma.^7-9 Consequently, IPV is underreported, particularly when injuries or presenting complaints do not immediately indicate violence. While specialized programs like Sexual Assault Nurse Examiners (SANE) offer in-hospital victim services, they are not universally available, and they primarily focus on sexual assault rather than the broader spectrum of IPV.¹⁰ Healthcare providers therefore rely on subtle clinical signals—such as recurring injury patterns and co-occurring health conditions—that may suggest IPV, even in the absence of direct disclosure.¹¹

Health informatics offers a promising approach to addressing this challenge by systematically identifying patterns in electronic health records (EHR) data that may be indicative of IPV. EHR data captures detailed information about patients’ injuries, diagnoses, and clinical histories during ED encounters, providing an opportunity to detect hidden associations between injury types, chronic conditions, and IPV risk. Advanced data mining techniques can uncover these associations, helping clinicians recognize IPV-related signals within routine clinical data.

However, ICD-10 diagnosis codes are not always accurately assigned, including those related to IPV and abuse.^{12, 13} As a result, some non-IPV patients may receive IPV-related diagnosis codes due to vague or ambiguous documentation. To address this, our study examined health condition patterns among confirmed IPV patients and those misclassified as IPV victims through incorrect ICD-10 code assignment. Understanding these differences in health condition patterns can help improve IPV identification and ensure more accurate diagnosis coding, ultimately supporting better clinical decision-making.

While prior research has explored IPV detection in healthcare data, much of this work has been narrowly focused on specific injury types (e.g., facial trauma) or limited to particular specialties (e.g., radiology or orthopedics).^14-17 In contrast, our study adopts a broader and more integrative approach, examining a wide range of physical and mental health conditions along with history of disease recorded in EHR data for female IPV survivors presenting to the ED. By combining qualitative and quantitative analyses, we aim to comprehensively characterize these patterns, offering a more holistic view of IPV-related clinical presentations.

1.1. Contributions

The contribution of this study is two-fold: First, we identified and evaluated a set of rules for detecting IPV-related patients that can be integrated into clinical practice and ED triage evaluation. We applied the Apriori algorithm to identify diagnostic patterns among IPV survivors and misclassified non-IPV patients, aiming to enhance clinicians’ recognition of IPV, even when undisclosed. Associations identified through Apriori are simple to interpret which makes them actionable in clinical practice. Second, we developed an annotated dataset of clinical notes from IPV-related patients. To support the Apriori analysis, we comprehensively reviewed and labeled ED encounters with gold-standard IPV and non-IPV classifications. This dataset provides a valuable foundation for advancing informatics approaches, informing data-driven screening tools, improving risk stratification, and supporting evidence-based clinical decision-making for timely intervention and care.

2. Methods

2.1. Data Source

Our study uses data from a ten-hospital academic health system in the mid-Atlantic region of the United States. The study cohort consists of female patients aged 18 years and older who presented to an emergency department (ED) within our healthcare system between January 1, 2022, and October 31, 2024. Demographic characteristics and ICD-10 diagnosis codes, which were recorded during the ED encounter, were extracted from the EHR database. Potential IPV victims were identified in a two-pronged approach. First, a list of IPV-explicit ICD-10 diagnosis codes (provided in Figure 1) obtained from a literature review were used to identify patients with IPV-explicit diagnosis codes documented during ED encounters.^18-21 Patients with any matching code were selected for chart review to confirm IPV status. Secondly, we obtained a separate list of abuse-related ICD-10 codes representing various forms of abuse (e.g., sexual, physical, psychological)^18-23 When using these codes, clinicians have the option to provide additional free-text context such as ‘abused by spouse on [DATE]’ to complement the standard ICD-10 description. For this analysis relating to IPV, we are interested in context containing the following partner-related keywords: marital, spouse, wife, husband, partner, domestic, or relationship problem. We use this combination of abuse-related ICD-10 codes and partner-related context to identify additional patients for our cohort. Figure 1 presents the approach to incorporate IPV-explicit andabuse-related diagnosiscodes to identify potential IPVencounters. Individuals withmore than one documented IPV encounter during the study period were removed from the analysis, ensuringthat only initial IPV encounters were considered. The study was approved by Institutional Review Board (IRB) at MedStar Health (IRB number = 0006202).

Figure 1.

The process to select potential IPV and non-IPV encounters for chart review.

2.2. Identification of IPV-related Encounters

To ensure accurate identification of IPV encounters, we conducted a chart review of 490 potential IPV-related encounters. Each encounter was manually reviewed, and unique encounters were classified as either IPV-related or non-IPV. For this study we incorporated the Centers for Disease Control and Prevention (CDC)’s definition of IPV, which defines IPV as abuse or aggression that occurs in a romantic relationship. An intimate partner refers to both current and former spouses and datingpartners. IPV can include physical or sexual violence, stalkingemotional abuse, and psychological aggression. Diagnosis codes recorded for these two cohorts were then compared using the Apriori algorithmto identify differences in health conditions between patients in abusive relationships and those misclassified as IPV patients by clinicians.

Potential IPV encounters identified through initial queries underwent further manual chart review. A two-person review team, consisting of one clinical reviewer and one reviewer with expertise in health equity, developed an initial codebook to guide the review process. Both reviewers had prior experience in chart review and qualitative analysis. Each reviewer independently reviewed all the clinical notes associated to the specific encounter and classified encounters as IPV or non-IPV. Chart reviews were conducted iteratively, with reviewers meeting after each round to reconcile disagreements and refine the codebook by adding new scenarios and de-identified examples.

All encounters were reviewed by the two reviewers with an agreement of 90.9% across all encounters. The IPV labeling disagreements were discussed with a subject matter expert who specializes in violence against women research. Disagreements in IPV classification primarily centered on encounters where patients presented with pain (e.g., chest pain) or mental healthconditions (e.g., severe anxiety) followingconflicts with their intimate partner. After consulting our subject matter expert, the reviewers determined that such encounters should be classified as IPV. Similarly, encounters involving suicidal ideation or attempts after intimate partner-related conflicts were labeled as IPV, given the established link between IPV and suicide risk.²⁴ However, encounters where the intimate relationship between the victim and perpetrator could not be discerned from clinical notes were labeled as non-IPV. Encounters where reviewers could not reach a consensus on the final classification, which accounted for 45 encounters representing 44 patients, were labeled as unclear and excluded from the cohort. Diagnosis codes from IPV and non-IPV encounters were analyzed to identify health condition patterns across the two cohorts.

2.3. Association Rule Mining

2.3.1. Feature Selection

ICD-10 diagnosis codes documented during ED encounters were analyzed to identify healthcondition patterns among IPV and non-IPVencounters. To prevent bias in the Apriori algorithm, IPV-explicit and abuse-relateddiagnosiscodes (listed in Figure 1), which were used to select encounters for chart review, were removed for all encounters, as these codes were present in every encounter.

Given the variability in ICD-10 coding for similar conditions (e.g., F41.1: Generalized anxiety disorder, F41.9: Anxiety disorder, unspecified, F41.8: Other specified anxiety disorders), all diagnosis codes were mapped to their associated Clinical Classifications Software Refined (CCSR) categories.²⁵ This mapping consolidates similar diagnosis codes into broader, clinically relevant categories, serving as a feature reduction approach. By reducing data dimensionality, CCSR categorization enhances the ability of association rule mining via Apriori algorithm to detect meaningful health condition patterns among confirmed IPV and non-IPV encounters. Analysis was conducted in Python 3.8. (Python Software Foundation).

2.3.2. Apriori Algorithm

This study applies the Apriori algorithm to identify associations between the diagnosis codes, which were mapped to their associated CCSR categories, documented during ED encounters for IPV and non-IPV patients to uncover patterns.^{26, 27} Apriori is a well-established, search-based data mining technique that identifies co-occurring items within large datasets, making it a promising approach for detecting clinically meaningful associations between physical injuries, mental health conditions, and history of diseases in this population.

Each rule generated by association rule mining takes the form LHS (left-hand side) → RHS (right-hand side), where in our analysis, LHS and RHS represent disjoint sets of ICD-10 diagnosis codes. The rule indicates that the presence of the LHS diagnosis codes increases the likelihood of the RHS diagnosis codes also being present. To assess the relevance and strength of each rule, we will calculate standard association rule mining metrics, including Support (the proportion of encounters containing the complete rule), Confidence (the conditional probability of RHS given LHS), and Lift (the observed co-occurrence relative to independence):

Support: This metric quantifies the proportion of IPV-related encounters that included a given diagnosis X. When looking at a combination of diagnoses, Support is a symmetrical measure, reflecting the joint occurrence of diagnoses X and Y within the same encounter.

Confidence: While Support is symmetrical, understanding the directional relationship between two diagnoses provides additional clinical insight. To capture this directionality, we applied Confidence. Confidence (X → Y) reflects the conditional probability of diagnosis Y occurring, given the presence of diagnosis X in IPV encounters. Confidence provides insight into the strength of the association between two diagnoses.

Lift: This is another key metric in association rule mining. Unlike Confidence, which can be biased toward frequent diagnoses, Lift adjusts for the baseline prevalence of the consequent diagnosis (Y). For example, if diagnosis Y appears in 80% of IPV-related encounters, any rule predicting Y would naturally have high Confidence, regardless of the true relationship between X and Y. Lift quantifies the strength of the association by scaling Confidence relative to the overall prevalence of Y and is calculated as Confidence (X → Y) divided by Support (Y). A Lift value greater than 1 indicates a positive association between X and Y, suggesting that diagnosis Y is more likely to occur when diagnosis X is present.

For the Apriori algorithm, the following minimum thresholds were established: a Support threshold of 0.1 to retain less frequent diagnoses, a Confidence threshold of 0.5 to identify stronger associations, and a Lift threshold of 1 to capture diagnoses that highly influence the occurrence of other diagnoses. This analysis focused on identifying health conditions with high confidence values, indicating strong associations between co-occurring conditions, as well as those with high Lift values, signifying a positive relationship between diagnoses.

3. Results

3.1. Cohorts’ Characteristics

There were 490 encounters associated with 474 female patients aged 18 years andolder who presentedto an EDwithin our healthcare system between January 1, 2022, and October 31, 2024 and had at least one IPV-explicit or abuse-related diagnosis code. Chart reviews confirmedthe IPV status for 378 (out of 490, 77.1%) encounters associated with 367 (out of 474, 77.4%) patients. Two patients had both IPV and non-IPV encounters. Those patients were dropped from the non-IPV encounters to prevent overlaps between the two cohorts. That lead to 67 non-IPV encounters representing 66 patients. A total of 14 patients with multiple IPV encounters and one patient with multiple non-IPV encounters were excluded, resulting in a cohort of 353 IPV and 65 non-IPV patients. Additionally, one IPV patient was excluded as their records contained only IPV-related diagnoses (listed in Figure 1), which were documented for all patients and could bias our results. The final study cohort consisted of 352 IPV and 65 non-IPV patients, each with one IPV or non-IPV encounters.

Table 1 summarizes the characteristics of the cohort. Among the 352 IPV patients, 322 (91.5%) were identified using IPV-explicit codes, while 30 (8.5%) were identified through abuse-related codes. Conversely, out of 65 non-IPV patients, all had IPV-explicit codes. The majority of IPV victims were single (65.3%) and Black (53.4%), with 32.4% aged 21–30 years and 62.2% covered by Medicare or Medicaid. All non-IPV encounters were assigned with Z63.0 which indicates problems in relationship with spouse or partner. Examples of excerpts from non-IPV encounters’ clinical notes are presented in the following:

Table 1.

Characteristics of patients confirmed as IPV and non-IPV via diagnosis codes.

	IPV N (% of 352)	Non-IPV N (% of 65)	Total N (% of 417)
ICD-10 Source
Identified through IPV-explicit diagnosis codes	322 (91.5)	65 (100.0)	387 (92.8)
Identified through abuse-related diagnosis codes	30 (8.5)	0	30 (7.2)
Patient Characteristics
Age
18-20	25 (7.1)	5 (7.7)	30 (7.2)
21-30	114 (32.4)	15 (23.1)	129 (30.9)
31-40	104 (29.5)	15 (23.1)	119 (28.5)
41-50	60 (17.0)	11 (16.9)	71 (17.0)
51+	49 (13.9)	19 (29.2)	68 (16.3)
Race
Black	188 (53.4)	25 (38.5)	213 (51.1)
White	118 (33.5)	25 (38.5)	143 (34.3)
Multiple	5 (1.4)	0	5 (1.2)
Other/Unknown	41 (11.6)	15 (23.1)	56 (13.4)
Insurance
Medicaid/Medicare	219 (62.2)	32 (49.2)	251 (60.2)
Commercial	88 (25)	27 (41.5)	115 (27.6)
Other/Unknown	45 (12.8)	6 (9.2)	51 (12.2)
Marital Status
Single	230 (65.3)	31 (47.7)	261 (62.6)
Married	88 (25)	28 (43.1)	116 (27.8)
Other	34 (9.7)	6 (4.6)	40 (9.6)

“…Patient states that she had a break-up with her first love and impulsively took four 800 mg ibuprofen…” “…Patient is feelingdepressed and anxious derivingfrommarital problemdue to infidelity issues by her husband. she has been feeling overwhelmed and came to the ED for possible med…”

Common themes among non-IPV patients included experiencing sadness or suicidal ideation following a recent breakup or partner infidelity, as well as having arguments with a partner without indications of abuse. Most of this group were single (47.4%) with 29.9% aged over 50 and 49.2% covered by Medicare or Medicaid. White and Black patients were equally represented (38.5% each).

3.2. Diagnosis Patterns Among IPV and non-IPV Cohorts

A total of 958 unique ICD-10 diagnosis codes were recorded for the IPV cohort, while 296 unique codes were documented for the non-IPV cohort. These diagnosis codes were mapped to their corresponding CCSR categories, resulting in 239 unique CCSR categories for the IPV cohort and 125 for the non-IPV cohort.

In the IPV cohort, the most frequent CCSR category was socioeconomic/psychosocial factors, appearing in 147 (41.8%) patients. Common diagnoses within this category included issues related to health literacy, legal circumstances, childhood physical or sexual abuse, unemployment, and homelessness. The second most frequent category was other specified status, documented for 130 (36.9%) of IPV patients. Frequent diagnoses within this category included long-term drug therapy and non-compliance with prescribed medication regimens. The third most common category was suicidal ideation/attempt/intentional self-harm, which was recorded for 108 (30.7%) patients in this cohort.

In non-IPV cohort, the most frequently recorded CCSR category was anxiety and fear-related disorders, found in 29 (44.6%) non-IPV patients. The second most frequent category was socioeconomic/psychosocial factors, recorded in 27 (41.5%) patients. Diagnosis codes commonly mappedto this category included problems relatedto primary support groups, disappearance or death of a family member, a history of childhood physical or sexual abuse, stressful life events affecting family and household, and family disruptions due to separation or divorce. The third most frequent CCSR category in this cohort was other specified status, recorded for 23 (35.4%) patients. Similar to the IPV cohort, frequent diagnoses mapped to this category included long-term drug therapy and non-compliance with prescribed medication regimens.

Table 2 presents the associations identified by Apriori algorithm which were recognized as clinically meaningful by our subject matter experts. Suicide ideation/attempt, socioeconomic/psychosocial factors, and mental health condition including depressive disorder, anxiety and fear-related disorders accounted for the majority of the item sets in the findings. These associations met the minimumthreshold criteria: a Support of at least 0.1, a Confidence of at least 0.5, and a Lift of at least 1.

Table 2.

The health condition patterns identified by Apriori algorithm. X and Y can represent one or more CCSR categories. The values are sorted by Lift.

CCSR Categories (X, Y)	Support (X)	Confidence (X→Y)	Lift (X→Y)	Category
((“socioeconomic/psychosocial factors”, “depressive disorders”), “suicidal ideation/attempt/intentional self-harm”)	0.10	0.82	2.67	IPV
((“depressive disorders”, “personal history of other disease”), “suicidal ideation/attempt/intentional self-harm”)	0.11	0.88	2.59	non-IPV
((“depressive disorders”, “personal history of other disease”), “suicidal ideation/attempt/intentional self-harm”)	0.11	0.79	2.58	IPV
(“sleep wake disorders”, (“socioeconomic/psychosocial factors”, “anxiety and fear-related disorders”))	0.11	0.58	2.53	non-IPV
(“depressive disorders”, “suicidal ideation/attempt/intentional self-harm”)	0.18	0.73	2.39	IPV
((“socioeconomic/psychosocial factors”, “personal history of other disease”), “suicidal ideation/attempt/intentional self-harm”)	0.11	0.72	2.35	IPV
((“depressive disorders”, ‘anxiety and fear-related disorders”), “suicidal ideation/attempt/intentional self-harm”)	0.14	0.75	2.22	non-IPV
((“depressive disorders”, “other specified status”), “suicidal ideation/attempt/intentional self-harm”)	0.12	0.73	2.15	non-IPV
((“depressive disorders”, “anxiety and fear-related disorders”), “personal history of other disease”)	0.11	0.58	2.11	non-IPV
(“superficial injury; contusion, initial encounter”, “other unspecified injury”)	0.12	0.53	2.08	IPV
(“depressive disorders”, “suicidal ideation/attempt/intentional self-harm”)	0.22	0.70	2.07	non-IPV
(“depressive disorders”, “anxiety and fear-related disorders”)	0.12	0.50	2.02	IPV
((“socioeconomic/psychosocial factors”, “anxiety and fear-related disorders”), “suicidal ideation/attempt/intentional self-harm”)	0.15	0.67	1.97	non-IPV
(“abnormal findings without diagnosis”, “socioeconomic/psychosocial factors”)	0.11	0.78	1.87	non-IPV
(“anxiety and fear-related disorders”, “suicidal ideation/attempt/intentional self-harm”)	0.14	0.56	1.84	IPV
(“sleep wake disorders”, “anxiety and fear-related disorders”)	0.14	0.75	1.68	non-IPV
((“depressive disorders”, “anxiety and fear-related disorders”), “other specified status”)	0.11	0.58	1.65	non-IPV
(“sleep wake disorders”, “socioeconomic/psychosocial factors”)	0.12	0.67	1.60	non-IPV
((“other specified status”, “anxiety and fear-related disorders”), “suicidal ideation/attempt/intentional self-harm”)	0.12	0.53	1.58	non-IPV
(“depressive disorders”, “other specified status”)	0.17	0.55	1.55	non-IPV
(“anxiety and fear-related disorders”, “suicidal ideation/attempt/intentional self-harm”)	0.23	0.52	1.53	non-IPV
(“anxiety and fear-related disorders”, “other specified status”)	0.14	0.55	1.49	IPV
(“anxiety and fear-related disorders”, “other specified status”)	0.23	0.52	1.46	non-IPV
(“suicidal ideation/attempt/intentional self-harm”, “other specified status”)	0.17	0.50	1.41	non-IPV
(“suicidal ideation/attempt/intentional self-harm”, “socioeconomic/psychosocial factors”)	0.18	0.58	1.40	IPV
(“depressive disorders”, “anxiety and fear-related disorders”)	0.18	0.60	1.34	non-IPV
(“anxiety and fear-related disorders”, “socioeconomic/psychosocial factors”)	0.14	0.55	1.32	IPV
(“other specified status”, “‘socioeconomic/psychosocial factors”)	0.20	0.54	1.29	IPV
((“other specified status”, “anxiety and fear-related disorders”), “socioeconomic/psychosocial factors”)	0.12	0.53	1.28	non-IPV
(“trauma- and stressor-related disorders”, “anxiety and fear-related disorders”)	0.12	0.57	1.28	non-IPV
(“anxiety and fear-related disorders”, “socioeconomic/psychosocial factors”)	0.23	0.52	1.25	non-IPV
(“depressive disorders”, “socioeconomic/psychosocial factors”)	0.13	0.51	1.23	IPV
(“suicidal ideation/attempt/intentional self-harm”, “socioeconomic/psychosocial factors”)	0.17	0.50	1.20	non-IPV

4. Discussion

4.1. Socioeconomics and Psychosocial Factors

No differences in demographic features were identified between the IPV and non-IPV cohorts. However, there were differences between CCSR categories. While both the IPV and non-IPV cohorts shared socioeconomic/psychosocial factors as the most frequent CCSR category, the diagnoses mapped to this category differed between the groups. In the IPV cohort, socioeconomic/psychosocial factorswere primarilyassociated with issues such as health literacy, legal circumstances, childhood physical or sexual abuse, unemployment, and homelessness. These factors reflect the living conditions of IPV survivors, where financial dependence and housing instability may exacerbate their vulnerability and dependencyon the abusive partner. Previousstudies have highlighted the intersection betweenfinancial instability and IPV, emphasizing how these factors contribute to the cycle of abuse.²⁸

Conversely, the non-IPV cohort’s socioeconomic/psychosocial factors were linked to problems related to primary support groups, the death or disappearance of a family member, childhood abuse history, stressful life events affecting the family and household, and family disruptions due to separation or divorce. These patterns appear to reflect general stressors and family dynamics rather than the direct impact of IPV, though the shared history of childhood abuse in both cohorts suggest a potential pathway for IPV later in life.²⁹ This overlap in early life trauma could explain why patients with such a history are sometimes misclassified as IPV survivors when diagnosis codes are assigned by ED clinicians.

4.2. Mental Health Conditions and IPV

The IPV cohort displayed strong associations between depressive disorders, personal history of other disorders, and physical injuries such as contusions. The personal history of other disorders in this cohort most commonly included suicidal behavior, a history of adult physical and sexual abuse, and non-suicidal self-harm. Previous research has established a well-documented link between IPV and suicidal ideation or attempts.²⁴ The presence of superficial injuries, particularly to the head and face, further supports thehypothesis that these patients are often victims of IPV.³⁰ Such injuries are typical of abuse victims, and their frequent occurrence alongside mental health conditions like depression and anxiety highlights the dual burden faced by IPV survivors.

In contrast, the non-IPV cohort showed strong associations between anxiety disorders, socioeconomic stressors, and sleep-wake disorders. Given the diagnoses mapped to the socioeconomic factors in this cohort—such as issues related to primary support groups, family disruptions, and stressful life events—it appears that the non-IPV cohort’s mental health issues are more likely linked to life stressors or family dynamics than to the direct consequences of IPV. Furthermore, the pattern”sleep-wake disorders, anxiety andfear-related disorders” was strongly present in this cohort, suggesting that these patients may be experiencing mental health conditions driven by stress from family circumstances, rather than IPV-related trauma.

It is important to note that the directionality inferred by association rule mining approaches is purely correlational. An implication X → Y, with Confidence c, simply means that c% of IPV-related encounters containing diagnosis X also containdiagnosis Y. Such a relationship should not be construed as implyingcausation without further analysis, which is typically beyond association rule mining and requires further experiments.

4.3. Clinical Implications

The findings from our association rule mining approach reveal critical health condition patterns that have important implications for EDclinicians. Specifically, the analysisuncovered a strongrelationshipbetween depressivedisorders, suicidal ideation, and housing and financial instability among patients who are survivors of IPV. Furthermore, strong associations were found between superficial injuries and other unspecified injuries, which are often seen in IPV cases. These patterns highlight the need for clinicians to be particularly vigilant when patients present with housing and financial instability, as it may be indicative of IPV. In such cases, clinicians should be prompted to screen for depressive disorders, suicidal ideation, and IPV, even if the patient does not directly disclose abuse. For IPV patients, these associations emphasize the importance of comprehensive care that addresses both physical and psychological trauma. Clinicians should be aware of the potential for trauma-related health conditions and ensure a thorough assessment that includes not only physical injuries but also mental health conditions and the patient’s living situation. This approach can help identify IPV more accurately, especially when the signs are subtle or when the patient is reluctant to disclose abuse.

On the other hand, for non-IPV patients, the presence of family-related stressors, anxiety, and other mental health issues suggests that these individuals may benefit from interventions focusing on external stressors such as social support and stress management. These factors may not be directly related to IPV but are important in addressing the patient’s overall well-being. The risk of misclassification, where non-IPV patients are mistakenly identified as IPV survivors, underscores the need for accurate diagnostic tools and refined screening methods. Misclassification can result in inappropriate interventions and a failure to address the actual underlying causes of the patient’s health issues.

Beyond supporting clinical decision-making, our findings have meaningful implications for improving patient-centered outcomes in emergency department settings. By leveraging latent diagnosis patterns indicative of IPV, clinicians may be better equipped to recognize and respond to abuse without relying solely on patient disclosure, which can often be hindered by fear, stigma, or emotional trauma. This proactive identification supports a more trauma-informed approachto care, potentially improvingtrust between patients and providers. Additionally, the ability to detect co-occurring mental health issues and socioeconomic stressors allows for more holistic interventions—such as timely referrals to social work, mental health services, or IPV advocacy—which can enhance patients’ engagement with care and improve their long-term health outcomes.

While automated Clinical Decision Support (CDS) tools based on EHRdata offer promise in enhancing IPVdetection, their deployment raises several ethical considerations. Privacy is a critical concern, especially given the sensitive nature of IPV and the risk of harm if information is mishandled or accessed without consent. Systems must include safeguardsto ensurethatIPV-relatedalertsare visible onlyto authorized providers andused within appropriate clinical contexts. There is also the risk of unintended stigmatization if patients are flagged incorrectly or prematurely labeled based on algorithmic patterns. To mitigate this, CDS tools should be designed to assist—not replace—clinical judgment, and providers should be trained in trauma-informed, nonjudgmental communication. Furthermore, transparency with patients about how their data is used, and giving them agency in their care decisions, is essential to maintaining trust. Ensuring equity in tool development and deployment is also critical, as algorithmic biases may disproportionately affect marginalized populations. Ethical deployment of IPV-related CDS must therefore balance predictive accuracy with sensitivity to the lived realities and rights of patients.

4.4. Limitations

The initial identification of the potential IPV cohort was based solely on IPV-related ICD-10 diagnosis codes. However, these codes are often underutilized in clinical documentation, meaning our sample may not fully represent all IPV patients who sought care in the ED. Incorporatingadditional EHRdata sources, suchas clinical notes, referrals to IPV-specific programs, and consultations, could help address this limitation. Additionally, we excluded patients with multiple IPV-related ED encounters during the study period to focus on identifying diagnosis patterns among new IPV cases. However, depending on the study timeline, some patients classified as new IPV victims may have previously visited the ED for IPV-related health consequences before the study window began in 2019. Extending the study period or integrating additional data sources, such as Chesapeake Regional Information System for our Patients (CRISP) data, could help mitigate this issue.

Although this study was conducted within a single academic health system in the mid-Atlantic region, the methods used—chart-reviewed classification, ICD-10 to CCSR mapping, and association rule mining—are broadly applicable to other healthcare systems with access to structured EHR data. However, generalizability of the specific diagnosis patterns we identified may be limited by regional differences in patient demographics, socioeconomic factors, and provider documentation practices. Academic centers may also differ from community hospitals in coding accuracy and resource availability. As such, replication of this methodology in other settings is necessary to validate and refine the rule sets. Future studies using multi-site or nationally representative datasets can further assess the transportability of these findings and support the development of generalized clinical decision support tools for IPV screening.

5. Conclusions

This study reveals critical differences and similarities in the health condition patterns between IPV patients and non-IPV patients. Our findings contribute to the development of informatics-driven screening algorithms, clinical decision support tools, and provider educationmodules that can contribute to more accuratecodingand facilitate moreeffective identification and intervention for this vulnerable patient population.