Homicides of Pregnant Women: Artificial Intelligence Detects Partner Violence and System Interaction

Abstract

Background:

Homicide ranks among the top causes of pregnancy-associated mortality in the United States. Intimate partner violence (IPV) has been implicated in violent maternal deaths, before which pregnant women may interact with health care, law enforcement, and legal systems.

Objective:

To understand IPV and system engagement prior to maternal deaths and to test the viability of using artificial intelligence (AI) for the analysis of narratives, we compared AI and human-rater analyses of National Violent Death Reporting System Restricted Access Data (NVDRS-RAD) narratives for IPV circumstances and system interactions.

Study Design:

We conducted a secondary data analysis of the female homicide records in the 2018–2020 NVDRS-RAD narratives. We trained a bidirectional encoder representations from transformers (BERT) model on 5,082 female nonpregnant cases, validating it with the 351 pregnant or recently pregnant cases. We conducted AI performance metrics for sensitivity, specificity, precision, and kappa values, identified key terms, and compared AI with human-rater analyses.

Results:

Among 351 complete NVDRS narrative records of pregnant or postpartum female homicide victims, 285 had primary suspects identified. Human-rater and AI analysis identified similar numbers for whether the suspect was a current or former partner and whether IPV history was noted before homicide. Natural language processing (NLP)-identified word patterns highlighted differences between IPV and non-IPV cases. Human raters identified 24% (80/351), compared with NLP’s identification of 21% (72/351), of pregnant women before death who interacted with health care and other systems. All AI models had strong performance metrics.

Conclusions:

Pregnant women in violent relationships interact with health care, law enforcement, and legal systems prior to their deaths. AI analysis is comparable with human raters in detecting IPV circumstances and system interactions among maternal homicides in the NVDRS. These findings highlight missed opportunities across sectors, underlining the importance of multisectoral interventions to prevent homicides of pregnant women.

Keywords

homicide intimate partner violence pregnant women prevention

Introduction

Despite being one of the wealthiest countries in the world, the United States faces the highest pregnancy-associated death rate among industrialized countries,¹ with 32.9 deaths per 100,000 live births in 2021.² Pregnancy-associated, but not related (PANR) mortality, as defined by the Maternal Mortality Review Committees and the Centers for Disease Control and Prevention (CDC), refers to deaths within 1 year of pregnancy from nonpregnancy related causes, such as homicide.³ Since the 2003 introduction of the pregnancy checkbox on death certificates, data suggest that trauma and injury have overtaken major pregnancy-associated medical complications (i.e., infection, bleeding, and high blood pressure) as leading causes of death of pregnant or postpartum women.⁴

Homicide, a preventable cause of PANR mortality, is a leading cause of death among the U.S. women under 45⁴ and is deeply intertwined with intimate partner violence (IPV). Nearly half of these homicides involve women killed by a current or former intimate partner,⁵ with nearly 15% of these victims being pregnant or recently postpartum. Although almost half of the pregnancy-associated homicides involve prior IPV incidents,⁶ the National Violent Death Reporting System (NVDRS) likely undercounts IPV-related deaths in PANR homicides due to limited IPV assessment.^4,7,8

Gaps in service response and research

Studies have shown that before homicide deaths, many female victims interacted with various systems. One study found that over 40% of victims killed by their partners obtained health care within a year of their deaths.⁹ In another study, 91% engaged with law enforcement (LE) within 3 years prior to being killed by an intimate partner, though of these cases, only 44% led to an arrest and 10% had sought legal protection orders.¹⁰ These interactions with health care, LE, and legal systems represent opportunities for intervention and prevention of femicide or women killed by their partners.

Despite known interactions of female homicide victims with health care, LE, and legal systems,^10,11 few studies have focused on pregnant and postpartum victims and their engagement with these systems before their violent deaths. One study reviewed the prenatal care records of 32 women murdered by a partner during the postpartum period and found that no other contact with systems was noted.¹² These findings underline the importance of IPV screening and follow-up during prenatal and postnatal health care encounters, as the American College of Obstetrics and Gynecology recommended.¹³ Similarly, although the research is limited, these women may have LE and legal interactions related to IPV, which could be opportunities to intercede. Investigating the circumstances and interactions is essential, and artificial intelligence (AI) has the potential to enhance these explorations significantly.

Machine learning in analysis of NVDRS narratives

Harnessing AI in the analysis of narratives is an evolving approach, and it holds the potential to assist in several ways. It offers enhanced accuracy in pattern recognition. It can also scale quickly, efficiently process large datasets, and enable broader trend analysis while handling incomplete and unstructured narratives better than traditional natural language processing (NLP) methods.^14,15 It also has the potential to minimize subjective human bias, although the risk of replicating human bias also exists.

Although the body of literature using AI NLP with NVDRS data is growing,^16–22 only three studies have employed AI NLP to analyze NVDRS narratives and IPV specifically, and these have focused on suicides.^23–25 Arseniev-Koehler et al. (2022) used predicted matching to investigate gender disparities in documenting mental health symptoms among suicide decedents. Kafka et al. (2023) deployed a supervised machine learning (SML) approach to analyze the co-occurrence of IPV and suicide for men and women. In 2024, Kafka et al. used SML and NLP to identify correlations between IPV and suicide, achieving significant model accuracy with a sensitivity of 0.70, specificity of 0.98, and precision of 0.72. This approach suggests the potential for machine learning to understand IPV circumstances in IPV-related homicides of pregnant and postpartum women. Within the small body of NVDRS literature using AI, none have focused specifically on homicides of pregnant and postpartum women.

Study aims

The purpose of this study was two-fold. First, we aimed to understand IPV and system engagement before maternal violent deaths using the NVDRS death narratives for women who were pregnant or recently pregnant. Second, we aimed to assess AI’s utility in identifying the occurrence of IPV prior to maternal deaths to ascertain if it could be a rapid and accurate approach for large datasets of narratives.

Materials and Methods

The study employed secondary data analysis using the NVDRS restricted access data (NVDRS-RAD) from 2018–2020. The NVDRS-RAD includes data on violent deaths, defined as “a death caused by the intentional use of physical force or power against a person, group, or community.” Data were obtained from death certificates, coroner and medical examiner (CME) records, and LE reports.²⁶ The CME and LE narratives are incident narratives that describe CME’s findings and summarize LE findings, respectively.²⁶ Data were collected from 48 states (excluding Florida and Hawaii), the District of Columbia, and Puerto Rico. Forty-six states had statewide data, while two states, California and Texas, offered data from selected counties representing a subset of their population (35 California counties, representing 71% of its population, and 4 Texas counties, representing 39% of its population). The District of Columbia and Puerto Rico had jurisdiction-wide data. Additional qualitative data from crime laboratories, hospitals, interviews, and IPV reports contributed to detailed narratives about the decedents’ and suspects’ demographic variables and incident circumstances.²⁶ These narratives supplement and validate NVDRS quantitative coded variables and help identify new and emerging risk factors relevant to violent deaths.

Sample

The study used deidentified data from the 2018–2020 NVDRS-RAD. A total of 17,597 homicide cases involving women of childbearing age (15–45) were available and reviewed for demographic context and narrative data quality (Table 1). Of these, 5,433 cases included structured and narrative data suitable for AI modeling—comprising 5,082 nonpregnant women and 351 women who were pregnant or had been within a year of their deaths (see Fig. 1). The NLP process used these narratives and selected categorical variables from the NVDRS dataset.

FIG. 1.

Sample sizes for training and testing natural language processing and conducting human-rater analysis. This figure illustrates the distribution of sample sizes used in the study for natural language processing (NLP) model training and testing and human-rater analysis.

Table 1.

Demographics and Narrative Completeness by Pregnancy Status

	Not pregnant	Pregnant/postpartum
Total number of cases	16,956	641
Cases with missing CME narrative	383	16
Cases with missing LE narrative	2,599	64
Average CME character count	820	825
Average LE character count	791	816
Average character count of narrative used per case	816	823
Mean age (years)	32.9 ± 9.7	27.7 ± 6.4
Race: White	11,905 (70.2)	330 (51.5)
Race: Black or African American	3,352 (19.8)	239 (37.3)
Race: Other/unknown	1,699 (10)	72 (11.2)
Ethnicity: Hispanic or Latino	2,006 (11.8)	74 (11.5)

CME, coroner and medical examiner; IPV, intimate partner violence; LE, law enforcement.

Data analysis

Human-Rater analysis of narratives

Two authors with expertise in qualitative methods, IPV, and forensic science reviewed all the narratives (both CME and LE) from the 351 female homicide victims who were pregnant or had been pregnant within a year of their deaths. The two raters examined the narratives to determine the frequency of IPV being present (inclusive of any combination of emotional, sexual, and/or physical violence mentioned in the narratives), the primary suspect being the partner or ex-partner, and engagement with LE, legal systems, and health care providers before their deaths. After analyzing the narratives independently, the findings were compared to evaluate interrater reliability. Overall, there was 95% agreement in the assignments of IPV mentioned in the narratives or not, primary suspect relationship to the victim, and different system contact. Any disagreements were resolved through discussion, with the first author making the final decision on the discrepancies. The human-rater analysis required several weeks to complete.

AI preprocessing and feature engineering

The data was preprocessed by cleaning, handling missing values, and encoding categorical variables for analysis. The categorical variables for analysis included manner of death (to isolate homicides from other causes), pregnancy status, IPV circumstance, intimate partner problem circumstance, and primary victim/suspect relationship. Categorical variables used for each prediction task are summarized in Table 2. Of note, NVDRS codes violence variables as “yes” and “no/not available/unknown,” making it more likely to underreport IPV and problems, given the potentially unknown nature of prior IPV in CME and LE narratives. Narrative texts were tokenized for the bidirectional encoder representations from transformers (BERT) model, with special tokens appended to signify the text’s start and end.^27,28 Stop words were retained to preserve the context while stemming or lemmatization was avoided to maintain the original form of words.²⁷ BERT extracted features from these token embeddings, and categorical variables were encoded and appended as additional features without employing dimensionality reduction, leveraging BERT’s capability to handle high-dimensional input efficiently. For the NLP analysis of IPV history, only cases with sufficient narrative detail and complete structured variables were included; records lacking adequate data were excluded to maintain model reliability. Structured NVDRS variables were appended to the narrative text as plain-language tags (e.g., “IntimatePartnerViolence: True”) and included in model training.

Table 2.

Categorical NVDRS Variables Used in BERT Model Training and Prediction Tasks

Prediction task	Variables used	Variable description
Primary suspect = Intimate partner	AbstractorDeathmanner_c	Manner of death (e.g., homicide = 2, suicide = 1)
	VicSusRelation	Victim-suspect relationship (e.g., partner, family, stranger)
	IntimatePartnerViolence_c	Indicates if IPV directly contributed to the death
	IntimatePartnerProblem_c	Relationship problems present
	CrisisIntimatePartnerProblem_c	Crisis due to relationship issues
	Pregnant	Victim was pregnant or postpartum
IPV history/relationship issues	Same as above (excluding death manner)	Same as above
System engagement: MH/HC	MentalHealthProblem_c	Victim had a diagnosed or suspected mental health condition
	CrisisMentalHealth_c	Acute mental health crisis prior to death
	MentalIllnessTreatment_c	Victim was in treatment for mental illness or substance use
	DisclosedtoHealthCareWorker	Victim disclosed distress or abuse to a health care provider
	CrisisPhysicalHealth	Physical health issue contributing to the fatal event
	Pregnant	Victim was pregnant or postpartum
System Engagement: Legal/LE	RecentCriminalLegalProblem_c	Legal/criminal issue recently affecting the victim
	LegalProblemOther_c	Other unspecified legal problems
	CrisisCivilLegal_c	Crisis related to civil legal matters
	HouseholdKnownToAuthorities_c	Household was previously known to law enforcement or child services
	CrisisCriminal_c	Criminal crisis contributed to the death
	Pregnant	Victim was pregnant or postpartum

All tasks used narrative text inputs from CME and LE reports in addition to the above structured variables.

BERT, bidirectional encoder representations from transformers; CME, coroner and medical examiner; IPV, intimate partner violence; LE, law enforcement; MH/HC, Mental Health/Health Care; NVDRS, National Violent Death Reporting System.

Model development

We employed the BERT uncased model, which is recognized for its robust text classification understanding of nuanced language and contextual relationships. This model was fine-tuned for three specific tasks: (1) Identifying if a current or former partner was the primary suspect in a homicide, using narratives and IPV-related, NVDRS quantitative variables; (2) detecting relationship issues from narrative texts and IPV-variables; (3) predicting system interactions, such as health care/mental health and legal/law enforcement, using both narratives and system-related NVDRS quantitative variables. The construct of “relationship issues” was derived from structured NVDRS variables (IntimatePartnerProblem, CrisisIntimatePartnerProblem, and IntimatePartnerViolence) for use in model training.

Models were trained using data from 5,082 nonpregnant female homicide victims and tested on 351 records of pregnant and postpartum female homicide victims. Hyperparameter tuning optimized learning rates, batch sizes, and training epochs for each task to ensure optimal performance. We used the Robustly Optimized BERT Pretraining Approach (RoBERTa) to support longer narrative inputs, and when narratives exceeded the token limit, chunking or truncation was applied to preserve critical content.

SHapley additive exPlanations

SHapley additive exPlanations (SHAP) values were used to interpret model predictions by highlighting influential keywords in the text.²⁷ Summary plots displayed the influence of these keywords. The top 10 influential keywords for each class in each task were extracted, showing the significant features driving the model’s decisions.

Statistical analysis

The models were tested on the full set of pregnant and postpartum cases, which had not been used during training. Model performance was evaluated using classification accuracy, precision, recall, and F1 score, with values closer to 1.0 indicating better accuracy (score range: 0–1.0).²⁹ Statistical significance testing validated the results, ensuring differences were not due to chance. The models were developed using Python (version 3.10) libraries such as Hugging Face Transformers (version 4.40.1),³⁰ Scikit-learn (version 1.2.2),³¹ and SHAP (version 0.45.0).²⁷ They were trained and evaluated on a cloud infrastructure with NVIDIA graphics processing units (L4 and T4), ensuring the computational resources necessary for fine-tuning BERT.³² The time involved in training and testing the AI NLP analyses was 30 hours.

Ethics and data privacy

This study followed ethical standards by ensuring data anonymity and utilizing secure data storage protocols. CDC approved the study design when access to the restricted data was granted; no additional consent was required. Data privacy was maintained by the CDC’s removal of identifiable information and the authors’ data encryption during analysis to protect sensitive information. The study received Institutional Review Board (IRB) approval from the first author’s university IRB.

Results

Prior intimate partner violence

Of the 351 female pregnant/postpartum victims, human raters found 191 (54%) experienced IPV before their deaths. Human raters identified common relationship issues among these women related to paternity disputes, desires to leave the relationship, accusations of infidelity, and prior domestic violence/IPV.

Intimate partner involvement in homicide

NLP analyses showed that not all female pregnant/postpartum victims had suspects noted in their case files; 81% (285/351) had an identified primary suspect. Of those, 65% (184/285) of primary suspects were current or ex-intimate partners. In cases in which the primary suspect was not a current or ex-partner, the women had been involved in drive-by shootings, armed robberies, gang-related shootings, or drug-related altercations.

System engagement before deaths

Human raters found that nearly one-fourth (24%, 80/351) of the pregnant/postpartum women who were killed had interactions with at least one system (health care, LE, legal) before their deaths (see Fig. 2). Of these 80 women who interacted with the systems before their death, 8 (10%) interacted with multiple systems. Human raters found that 50 women (62.5%) had law enforcement or legal system interactions only, such as previous wellness checks at the residence, prior domestic dispute calls, and restraining and/or protective orders filed. In addition, 22 (27.5%) women interacted with health care or mental health services only, separate from the prenatal care they may have received. Human raters noted interaction temporality, some within a day of the fatal incident; NLP analyses did not identify temporality.

FIG. 2.

Pregnant/postpartum women with system interactions before their deaths (n = 80). This Venn diagram illustrates the overlap in system interactions documented in the coroner/medical examiner and law enforcement narratives for pregnant and postpartum homicide victims.

NLP accuracy

Primary suspect as intimate partner

After training the NLP model in the nonpregnant female homicide dataset (n = 5,082), we tested the model on the pregnant/postpartum set (n = 351). The test on pregnant/postpartum victims emulated the human-coded data, finding that 285 women had a primary suspect identified. NLP further identified 170/285 (60%) women with a partner/ex-partner as the primary suspect, 14 records fewer than the human raters. NLP results are shown in Table 3, including the model’s high precision, recall, and accuracy. Top influential keywords for partner involvement included “violence,” “estranged,” “divorce,” and “boyfriend,” while keywords for other suspects are “retaliation,” “speculation,” “witnesses,” and “authorities.” Of note, the most influential word for those experiencing IPV was “sealed” reflecting instances where the CME report was “sealed by the court,” but the LE report was not. In all cases of a CME report being “sealed by the court,” relationship issues or violence were involved, and the primary suspect was the partner or ex-partner. These keywords helped in the model’s predictions. The additional women (n = 14) that human raters identified included narrative phrases such as “dating” and “sugar daddy,” which may not have been flagged by NLP.

Table 3.

Partner/Ex-Partner as Primary Suspect in Homicides of Pregnant or Postpartum Women

NLP model results	Precision	Recall	F1-score	Support
Partner/ex-partner, not primary suspect	1.00	0.94	0.97	115
Partner/ex-partner primary suspect	0.96	1.00	0.98	170
Accuracy			0.98	285
Macro average	0.98	0.97	0.97	285
Weighted average	0.98	0.98	0.98	285

Precision: how many of the “positive” predictions made by the model were correct; Recall: how many actual positives were classified correctly as positives; Accuracy: frequency of the model making a correct prediction across the entire dataset; F-1 score: combines precision and recall using their harmonic mean, maximizing both precision and recall.

Relationship issues, including past IPV

The results of NLP analyses identifying relationship problems in narratives are shown in Table 4, indicating good precision (67% and 90%), recall (83% and 79%), and accuracy (80%). NLP identified 176 women having past IPV, resulting in an IPV history detection rate of 66% (176/266) within the smaller sample. Figure 3 conveys the top 10 influential words for identifying narratives of women with relationship issues, while Figure 4 shows the influential words for identifying narratives of women without relationship issues.

FIG. 3.

SHAP values for top 10 influential keywords in narratives with relationship problems in homicides of pregnant and postpartum women. This figure displays the SHapley Additive exPlanations (SHAP) values, which range from 0 to 1, and highlight the top 10 influential keywords in narratives where relationship issues, including intimate partner violence (IPV), were identified. Higher SHAP values indicate a stronger influence of the keyword in the model’s prediction of relationship problems.

FIG. 4:

SHAP values for top 10 influential keywords in narratives without relationship problems in homicides of pregnant and postpartum women. This figure displays the SHapley Additive exPlanations (SHAP values, which range from 0 to 1, and highlights the top 10 influential keywords in narratives without relationship problems. Higher SHAP values indicate a stronger influence of the keyword in the model’s predictions.

Table 4.

Relationship Issues Prior to Homicides of Pregnant or Postpartum Women

NLP model results	Precision	Recall	F1-score	Support
No relationship issues	0.67	0.83	0.74	90
Relationship issues, conflict even if not directly related to homicide	0.90	0.79	0.84	176
Accuracy			0.80	266
Macro average	0.79	0.81	0.79	266
Weighted average	0.82	0.80	0.81	266

Prior interactions with LE, legal, and mental health services

The analysis of narratives to identify system interactions yielded an NLP model that accurately identified these interactions based on AI performance metrics (Table 5). Still, the model was not as accurate as human raters. The NLP identified only 7% (26/351) interacting with mental health services only, 3% (10/351) with LE or legal systems only, and 10% (36/351) with any combination of these systems.

Table 5.

System Interactions Prior to Homicides of Pregnant and Postpartum Women

NLP model results	Precision	Recall	F1-score	Support
Mental health/health care services	0.54	0.77	0.63	26
Law enforcement/legal services	0.15	0.40	0.22	10
None	0.97	0.89	0.93	315
Micro average	0.86	0.87	0.87	351
Macro average	0.55	0.69	0.60	351
Weighted average	0.91	0.87	0.89	351
Samples average	0.87	0.87	0.87	351

Discussion

Intimate partner violence and homicides

Homicide is a leading cause of death of pregnant women in the United States,³³ which can explain both human-rater and AI analyses detecting high rates of IPV among homicides of pregnant and postpartum women as documented in NVDRS narratives. Our findings support the prevalence of IPV in fatalities of women, with over half of the pregnant and postpartum women included in this study experiencing IPV prior to their deaths. However, this figure may underestimate the true prevalence, as there is likely an overreporting of “unknown” IPV presence when authorities complete CME and LE reports.

System interactions

One in four pregnant or postpartum women in this study had interactions with LE, legal, mental health, or other health care services prior to their deaths. This differs from Kafka et al.’s findings (2024) in Washington state, where nearly 50% of men and women of all ages had some prior system contact before their deaths.³⁴ Although our sample was specific to pregnant and postpartum women across the entire United States, the prevalence difference is notable and worth further exploration. Women experiencing IPV can have their access to health care controlled by their partner,³⁵ and this may be more prevalent during and after pregnancy. Women experiencing abuse may also fear for their safety if they present to health care and disclose IPV in states with health care providers’ mandatory reporting to LE of patient injuries due to violence,³⁶ and this can be an elevated concern when an expectant or new mother is being seen. These factors likely lead to low detection rates of system interactions. Moreover, our study found fewer system interactions than previous non-NVDRS research.^10,11 Because information on system interactions, including the presence of prenatal care, is not always noted in the narratives, the system interaction rate we found is likely underestimated. Health care interactions are certainly underestimated, given that the CDC estimates >85% of women receive at least some prenatal care.³⁷

Clinical implications

To prevent potential homicides, clinicians may interpret these findings in several ways. First, the sheer number of pregnant and postpartum women experiencing fatal IPV reinforces the need for IPV screening, not only for pregnant women but also for women who have recently been pregnant. In addition, clinicians should be trained in how to assess for homicide risk and help clients with safety planning during pregnancy.³⁸ Additionally, for more accurate postmortem data collection, coroners and medical examiners should consistently document both the pregnancy status of women and any signs of trauma (i.e., bruising patterns, etc.) to assist researchers in characterizing the scope of the problem. The lack of CME/LE documentation of health care encounters, including prenatal care, likely underrepresents their true interactions. NVDRS narratives and our findings reinforce the importance of multisectoral awareness and strengthening partnerships between LE, legal, and health care systems. Compared with human raters, the time saving in training NLP models on detecting IPV and system interactions among medical records or other narratives could help health care providers, medical examiners, LE, and other legal staff screen women for IPV in their settings, or at least more quickly understand the magnitude of the problem and missed opportunities for prevention and intervention.

NLP efficacy

In this study, we had mixed results in the efficacy of AI-driven compared with human raters’ analysis to detect IPV circumstances and system interactions among maternal homicides in the NVDRS. The BERT models replicated the identification of primary suspects well. Still, the models did not replicate findings for primary suspects who were intimate or ex-intimate partners at the same level of accuracy. Identifying IPV in relationships was relatively accurate, albeit with a smaller sample. The system interactions model did not replicate human-rater analysis perfectly, partly because human raters could identify nuances in language. Transformer-based models like BERT consider word context, not just keyword presence, so while SHAP highlights influential terms, it is surrounding context that drives predictions.

AI NLP offers several potential further applications. It could be integrated with raw data sources, such as medical, legal, and law enforcement records (such as electronic health records or court documents), to assess an individual’s risk and potentially to intercede and prevent intimate partner homicide comprehensively. Future research could examine the effectiveness of embedding a trained AI program in clinical or law enforcement systems to flag and address high-risk cases. Because specific AI improves over time with new data, the approach will stay relevant as language and risk factors evolve.

Methodological implications

For most women in our study, pre- and postnatal care was not captured, as there are no variables in the dataset or common language in the CME/LE reports indicating these health care encounters. Generating standard terms used in both CME and LE narratives to describe pregnancy, as well as partner and ex-partner violence and system interactions, would assist in consistency for abstraction and data analysis. In addition, future research and clinical work can consider if applying AI and NLP models directly to the text data in pregnant patients’ electronic medical records could help detect IPV and thereby notify the clinicians of the need for formal homicide risk assessment screening.

Human-rater analysis of narratives for system interactions identified some aspects of the system interactions that were not assessed in the NLP approach, including the timing of the interaction relative to the homicide. A model may be trained to search for time-frame text. Still, this training would require a larger sample or consistent record keeping of the temporal nature of the system interactions, as the number of records with this information was relatively small in our current sample.

Our study highlights the potential of deep learning for analyzing narratives with good accuracy. The sample size was relatively small for NLP purposes, which may have accounted for human-rater analysis identifying more IPV cases and system interactions than the NLP models. As it was, the NLP method halved the time required for narrative analysis. If the dataset had included many more narratives, the time required for human raters to review records accurately would have increased exponentially. In contrast, deep learning models can efficiently scale to larger datasets with minimal additional processing time, and their accuracy could improve with more training data.

Strengths and limitations

This study was the first of its kind to use a near-national sample of pregnant and postpartum female homicides to assess IPV and system interactions and the first to compare AI and human rater analyses of IPV and system interactions for pregnant and postpartum women. As such, the study has several strengths. By restricting the dataset to homicides of pregnant or postpartum women, we characterize maternal mortality due to homicide. Using only homicides, we could identify more explicit patterns of IPV and system interactions for the AI model to learn and detect relevant circumstances. If the analysis had included nonhomicide deaths (e.g., accidental deaths, suicides), it could have introduced “noise” or unrelated text, leading to irrelevant patterns that might have confused the AI model.

At the same time, there are several limitations. The analysis could only look at completed homicides and not attempted killings. As such, we may miss potential system interactions or IPV scenarios where pregnant or postpartum women were at risk but not killed. For example, women experiencing IPV might interact with health care or LE but survive, and these interactions could offer valuable insights into missed intervention opportunities. In addition, given there were only 351 pregnant or postpartum women, this may have potentially affected the ability to transfer learning from the training to the testing, as the small number of pregnancy-related homicides might limit the AI’s ability to generalize well to this specific population. As SHAP values were calculated on the full narrative input (including appended NVDRS tags), the individual contribution of structured variables was not isolated, highlighting a need for future refinement.

The NVDRS data have limitations that could lead to underestimating the true prevalence of health care engagement among pregnant and postpartum women. Notably, while many individuals in these groups likely interacted with maternity care providers, given nearly 75% of U.S. pregnant women begin prenatal care in the first trimester,³⁹ such interactions were not captured in our dataset. Additionally, the pregnancy/postpartum status of decedents may be underreported, which raises concerns about potential misclassification. Future iterations of the NVDRS may identify further cases not captured through the current approach. Despite these constraints, our study provides a critical initial assessment, and we emphasize the need for ongoing refinement of methodologies to enhance the accuracy of maternal health-related data.

Furthermore, although NVDRS narratives serve as a valuable resource for advancing research on violent maternal deaths, they are not without potential biases and challenges, specifically related to data collection at the state level (missing CME/LE narratives for some, missing pregnancy status, missed IPV coding) and abstraction across all records at the CDC. These arise from a still-fragmented U.S. death investigation system.⁴⁰ The NVDRS itself recognizes this limitation.

Conclusions

IPV lethality is of grave concern, as homicides of women who are pregnant or recently pregnant have widespread ramifications on society. AI-driven analysis is a significant time-saver, and when the sample size is large enough, the results can be very accurate. Using AI to assist in analysis can support future research to inform multisectoral interventions aimed at preventing tragic deaths of vulnerable women at the hands of their partners. Addressing the preventable causes of maternal mortality can improve familial well-being, productive capacity in women of reproductive age, and the economic well-being of families and communities.

Footnotes

Acknowledgments

This research uses data from NVDRS, a surveillance system designed by the Centers for Disease Control and Prevention's (CDC) National Center for Injury Prevention and Control. The findings are based, in part, on the contributions of the funded states/territories/jurisdictions that collected violent death data and the contributions of their partners, including personnel from law enforcement, vital records, medical examiners/coroners, and crime laboratories. The analyses, results, and conclusions presented here represent those of the authors and not necessarily reflect those of CDC.

Author Disclosure Statement

None of the authors have competing interests to declare.

Funding Information

This work was supported by the Agency for Healthcare Research and Quality (AHRQ) through the Center for Immersive Learning and Digital Innovation: A Patient Safety Learning Laboratory advancing patient safety through design, systems engineering, and health services research (Award Number 7R18HS029124-03; PI: V.P.). The content is solely the responsibility of the authors and does not necessarily represent the official views of the AHRQ.

Data Statement

Persons interested in obtaining data files from NVDRS-RAD should contact CDC’s National Center for Injury Prevention and Control, 4770 Buford Hwy, NE, MS F-64, Atlanta, GA 30341-3717, (800) CDC-INFO (232-4636).

References

Credit Suisse Research Institute. The Global Wealth Report 2022. 2022. Available from: https://www.credit-suisse.com/about-us/en/reports-research/global-wealth-report.html

Hoyert

. Maternal Mortality Rates in the United States, 2020. National Center for Health Statistics. 2022.

Maternal Mortality Review Committees (MMRC). Review to action: Definitions. Available from: https://reviewtoaction.org/learn/definitions [Last accessed: March 20, 2023].

Campbell

, Matoff-Stepp

, Velez

, et al. Pregnancy-associated deaths from homicide, suicide, and drug overdose: Review of research and the intersection with intimate partner violence. J Womens Health (Larchmt), 2021; 30(2):236–244; doi: 10.1089/jwh.2020.8875

Petrosky

, Blair

, Betz

, et al. Racial and ethnic differences in homicides of adult women and the role of intimate partner violence - United States, 2003–2014. MMWR Morb Mortal Wkly Rep, 2017; 66(28):741–746; doi: 10.15585/mmwr.mm6628a1

Palladino

, Singh

, Campbell

, et al. Homicide and suicide during the perinatal period: Findings from the National Violent Death Reporting System. Obstet Gynecol, 2011; 118(5):1056–1063; doi: 10.1097/AOG.0b013e31823294da

Campbell

, Webster

, Koziol-McLain

, et al. Assessing risk factors for intimate partner homicide. National Institute of Justice Journal, 2003; 250(11):14–19.

Campbell

, Glass

, Sharps

, et al. Intimate partner homicide: Review and implications of research and policy. Trauma Violence Abuse, 2007; 8(3):246–269; doi: 10.1177/1524838007303505

Sharps

, Koziol-McLain

, Campbell

, et al. Health care providers’ missed opportunities for preventing femicide. Prev Med /2001; 33(5):373–380; doi: 10.1006/pmed.2001.0902

10.

Koppa

, Messing

. Can justice system interventions prevent intimate partner homicide? An analysis of rates of help seeking prior to fatality. J Interpers Violence, 2021; 36(17–18):8792–8816; doi: 10.1177/0886260519851179

11.

O’Neill

, Buajitti

, Donnelly

, et al. Characterising mental health and addictions and assault-related health care use in the year prior to death: A population-based linked cohort study of homicide victims. Int J Popul Data Sci, 2021; 6(1):1410; doi: 10.23889/ijpds.v6i1.1410

12.

Cheng

, Horon

. Intimate-partner homicide among pregnant and postpartum women. Obstet Gynecol, 2010; 115(6):1181–1186; doi: 10.1097/AOG.0b013e3181de0194

13.

ACOG Committee Opinion No. 518: Intimate partner violence. Obstet Gynecol, 2012; 119(2 Pt 1):412–417; doi: 10.1097/AOG.0b013e318249ff74

14.

Supriyono , Wibawa

, Suyono , et al. Advancements in natural language processing: Implications, challenges, and future directions. Telematics and Informatics Reports, 2024; 16:100173; doi: 10.1016/j.teler.2024.100173

15.

Leeson

, Resnick

, Alexander

, et al. Natural language processing (NLP) in qualitative public health research: A proof of concept study. Int J Qual Methods, 2019; 18; doi: 10.1177/1609406919887021

16.

Wang

, Dang

, Sun

, et al. An NLP approach to identify SDoH-related circumstance and suicide crisis from death investigation narratives. J Am Med Inform Assoc, 2023; 30(8):1408–1417; doi: 10.1093/jamia/ocad068

17.

Wang

, Zhou

, Han

, et al. A natural language processing approach to detect inconsistencies in death investigation notes attributing suicide circumstances. Commun Med (Lond), 2024; 4(1):199; doi: 10.1038/s43856-024-00631-7

18.

Arseniev-Koehler

, Foster

, Mays

, et al. Aggression, escalation, and other latent themes in legal intervention deaths of non-hispanic black and white men: Results from the 2003–2017 National Violent Death Reporting System. Am J Public Health, 2021; 111(S2):S107–S115; doi: 10.2105/AJPH.2021.306312

19.

Arseniev-Koehler

, Cochran

, Mays

, et al. Integrating topic modeling and word embedding to characterize violent deaths. Proc Natl Acad Sci U S A, 2022; 119(10):e2108801119; doi: 10.1073/pnas.2108801119

20.

Mezuk

, Ko

, Kalesnikava

, et al. Suicide among older adults living in or transitioning to residential long-term care, 2003 to 2015. JAMA Netw Open, 2019; 2(6):e195627; doi: 10.1001/jamanetworkopen.2019.5627

21.

, Kalesnikava

, Jurgens

, et al. A data science approach to estimating the frequency of driving cessation associated suicide in the US: Evidence from the national violent death reporting system. Front Public Health, 2021; 9:689967; doi: 10.3389/fpubh.2021.689967

22.

Lohman

, Ko

, Rapp

, et al. State variation in long-term care availability, regulation, and cost and suicide mortality among older adults in the United States: 2010-2015. J Am Med Dir Assoc, 2021; 22(11):2337–2343.e3; doi: 10.1016/j.jamda.2021.02.008

23.

Kafka

, Fliss

, Trangenstein

, et al. Detecting intimate partner violence circumstance for suicide: Development and validation of a tool using natural language processing and supervised machine learning in the National Violent Death Reporting System. Inj Prev, 2023; 29(2):134–141; doi: 10.1136/ip-2022-044662

24.

Kafka

, Moracco

, Pence

, et al. Intimate partner violence and suicide mortality: A cross-sectional study using machine learning and natural language processing of suicide data from 43 states. Inj Prev, 2024; 30(2):125–131; doi: 10.1136/ip-2023-044976

25.

Arseniev-Koehler

, Mays

, Foster

, et al. Gendered patterns in manifest and latent mental health indicators among suicide decedents: 2003-2020 National Violent Death Reporting System (NVDRS). Am J Public Health, 2024; 114(S3):S268–S277; doi: 10.2105/AJPH.2023.307427

26.

National Violent Death Reporting System Web Coding Manual, Version 6 (Centers for Disease Control and Prevention) (2022).

27.

Lundberg

, Lee

. A unified approach to interpreting model predictions. NIPS, 2017.

28.

Khurana

, Koli

, Khatter

, et al. Natural language processing: State of the art, current trends and challenges. Multimed Tools Appl, 2023; 82(3):3713–3744; doi: 10.1007/s11042-022-13428-4

29.

Hicks

, Strümke

, Thambawita

, et al. On evaluation metrics for medical applications of artificial intelligence. Sci Rep, 2022; 12(1):5979; doi: 10.1038/s41598-022-09954-8

30.

Wolf

, Debut

, Sanh

, et al. Transformers: State-of-the-Art Natural Language Processing. Hugging Face; 2020.

31.

Pedregosa

, Varoquaux

, Gramfort

, et al. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 2011; 12:2825–2830.

32.

NVIDIA. Fine-tune and optimize BERT: Jupiter Notebooks for BERT pre-training, fine-tuning, and inference profiling and optimization. NVIDIA. Available from: https://catalog.ngc.nvidia.com/orgs/nvidia/containers/bert_workshop?ncid=em-nurt-245273-vt33 [Last accessed: August 15, 2024].

33.

Wallace

, Gillispie-Bell

, Cruz

, et al. Homicide during pregnancy and the postpartum period in the United States, 2018–2019. Obstet Gynecol, 2021; 138(5):762–769; doi: 10.1097/AOG.0000000000004567

34.

Kafka

, Adhia

, Martin

, et al. Deaths related to domestic violence in Washington State. JAMA Netw Open, 2024; 7(9):e2429974; doi: 10.1001/jamanetworkopen.2024.29974

35.

Miller

, McCaw

. Intimate partner violence. N Engl J Med, 2019; 380(9):850–857; doi: 10.1056/NEJMra1807166

36.

Sullivan

, Hagen

. Survivors’ opinions about mandatory reporting of domestic violence and sexual assault by medical professionals. Affilia, 2005; 20(3):346–361; doi: 10.1177/0886109905277611

37.

Martin

, Osterman

MJK

. Changes in prenatal care utilization: United States, 2019–2021. Natl Vital Stat Rep, 2023; 72(4):1–14.

38.

Messing

, Campbell

, Snider

. Validation and adaptation of the danger assessment-5: A brief intimate partner violence risk assessment. J Adv Nurs, 2017; 73(12):3220–3230; doi: 10.1111/jan.13459

39.

Martin

, Hamilton

, Osterman

MJK

. Births in the United States, 2023. NCHS Data Brief, no 507. Hyattsville, MD: National Center for Health Statistics. 2024; doi: 10.15620/cdc/158789

40.

Dang

, Kahsay

, James

, et al. Research utility and limitations of textual data in the National Violent Death Reporting System: A scoping review and recommendations. Inj Epidemiol, 2023; 10(1):23; doi: 10.1186/s40621-023-00433-w