Article Text

Download PDFPDF
Making the ‘invisible’ visible: transforming the detection of intimate partner violence
  1. Bharti Khurana1,
  2. Steven E Seltzer1,
  3. Isaac S Kohane2,
  4. Giles W Boland1
  1. 1 Department of Radiology, Brigham and Women's Hospital, Boston, Massachusetts, USA
  2. 2 Department of Bioinformatics, Harvard Medical School, Boston, Massachusetts, USA
  1. Correspondence to Dr Bharti Khurana, Department of Radiology, Brigham and Women's Hospital, 75 Francis St, Boston, MA 02115, USA; bkhurana{at}

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.


On 25 November 2018, the United Nations chillingly reported that the most dangerous place for women is inside their own homes. Each year more than half of female homicides are committed by current or former intimate partners or family members.1 Intimate partner violence (IPV), within the domestic violence spectrum, is defined as physical, sexual or emotional violence between partners or former partners.2 It is a serious public health concern with millions of people experiencing violence at the hands of an intimate partner. WHO recognizes IPV as a global issue, prevalent at epidemic proportions in every society, socioeconomic and educational group. According to the National Intimate Partner and Sexual Violence Survey, one in four women and one in nine men in USA have reported severe form of physical violence by an intimate partner during their lifetime.3 Despite the high prevalence and urgency of this critical public health issue, IPV continues to be profoundly underdiagnosed and is considered a persistent hidden epidemic. In addition to physical injuries, IPV has both short-term and long-term negative health consequences including asthma, irritable bowel syndrome, diabetes, poor reproductive health, chronic pain syndrome and mental health problems.4 With victims of IPV seeking medical care more often, healthcare providers can play a vital role in reducing the devastating impact of IPV by representing a trusting source of divulging abuse. The major obstacle to its early detection and intervention is victim under-reporting of physical violence to healthcare providers. Screening for IPV can be an effective tool for detecting and preventing future violence. However, several barriers limit the use and success of these screening programs. Due to shame, privacy, economic dependency, fear of retaliation, legal factors or lack of trust of providers, a patient may not self-report and even fabricate the history of her injury.5 The Center for Disease Control has also reported several barriers for providers in administering questions including fear of offending the patient or partner, time constraints, discomfort with the topic, need for privacy, perceived lack of power to change the problem and a misconception regarding patient’s population risk of exposure to IPV.6 As a result, the proportion of identifiable IPV cases to date only represents the tip of the iceberg while an overwhelming number of IPV cases go undiagnosed with consequently missed opportunities to offer timely preventive services. IPV is also associated with profound economic costs with one study reporting healthcare costs of approximately $19.3 million, each year, for every 1000 000 women (age range, 18-64 years).7 Clearly, improvements in the detection and appropriate triage are urgently needed to address this major public health issue.

Imaging has made substantial contributions to the detection of non-accidental trauma in children.8 Laws now mandate the detection and reporting of even the potential of a non-accidental trauma in children, so identification of paediatric violence is an essential component of a radiologist’s training. Sadly, for adult victims, the situation is very different. The role that imaging could play to influence the detection of adult victims, similar to paediatric patients, has never been explored until recently.9 10 Radiologists are trained to simply report adult traumatic findings from the current examination without making any active effort to highlight any possibility of this life-threatening issue. Complicating matters, IPV victims often present to disparate healthcare providers in different healthcare settings, convoluting any unified efforts to discover patients being harmed.

The rapid advances in imaging, information technology and machine learning, an application of artificial intelligence (AI) that provides automated learning, present an opportunity to achieve a breakthrough in identifying the victims of IPV and thereby activating timely intervention. This article describes how imaging data and its integration with robust AI-enabled tools can generate predictive information for patients at risk, who may not otherwise be forthcoming, thereby empowering providers to intervene.

Transforming the role of radiology

The key is to recognise that a certain pattern and distribution of injuries present on radiological studies correlate to IPV (figure 1). The location of injuries such as mid-face especially nasal, zygomatic, periorbital, mandibular angle/condyle injuries is commonly reported in IPV victims.11 In addition to concussions, non-fatal strangulation is a common mode of assault with IPV that can result in recurrent traumatic brain injuries and heightens the risk of morbidity and mortality among the patients experiencing IPV. These victims may have no physically visible injuries and may present days or weeks later with a variety of neurological symptoms including headaches, confusion, vertigo, amnesia, and loss of consciousness due to combined interruption of blood flow and hypoxia.12 Such symptoms often require patients to undergo head, face, neck, cervical spine CT or CT angiography (CTA). Mid-facial anatomy, the key area of interest in IPV, is comprehensively imaged by any one of these studies. Injury to laryngeal cartilages is also visible on neck CT and neck CTA performed to assess vascular dissection. Furthermore, thoracic injuries are common in IPV, which may present with chest pain and/or shortness of breath. Acute and/or chronic posterior rib, clavicular or scapular fractures can be identified on chest radiographs or chest CT performed either for pulmonary embolism or aortic dissection.

Figure 1

Transforming the role of radiology in intimate partner violence (IPV) detection by recognising high imaging utilisation, injury location and patterns specific to IPV, injuries’ inconsistent history and old injuries of different body parts.

A defensive injury pattern, one where the victim is attempting to avert and/or minimise the blow, can be a critical imaging correlate of IPV.13 Figure 2 demonstrates a non-displaced ulnar fracture in a 38-year-old woman presenting to the emergency department with a self-reported history of fall. However, an ulnar fracture resulting from such a fall is very unusual as most of the force at impact is transmitted to the distal radius, which may fracture as a result. An ulnar fracture, on the other hand, is often sustained when a forearm is raised in a defensive manner to protect the face or head. The patient later shared that her boyfriend had hit her during an argument. Similarly, contusions or other fractures (as listed above) inconsistent with the provided history are suspicious for IPV.

Figure 2

PA radiograph of the wrist in a 38-year-old woman with a reported history of fall reveals a non-displaced fracture of the distal ulna (white arrow). PA, Posteroanterior.

The simultaneous presence of more than one radiological finding of injury and/or the existence of multiple fractures at various stages of healing increase the likelihood that the patient’s injuries have resulted from IPV. For example, a patient with both a distal ulnar and nasal bone fracture would increase the probability of IPV, regardless of whether these injuries occurred concurrently or at different times. Radiologists and providers must therefore consistently be on the alert and must mine the electronic medical records and image archiving systems to review prior imaging studies of different body parts to elucidate any hint of IPV, which can then be correlated with clinical findings and the presenting history.

High imaging utilisation has also been reported among the victims of IPV.10 For instance, a young patient with multiple facial CTs and upper extremity radiographs should alert the radiologist or provider to suspect recurrent violence. These studies could have been performed at different institutions. With increasing healthcare consolidation, providers may now have access to the electronic health records (EHR) from different institutions, necessary to cross-correlate imaging and other clinical findings, increasing the chance of identifying a presenting patient with IPV, who might otherwise fail to disclose her prior injuries. Due to current technical and legal constraints, it is often challenging to access imaging performed outside the health network. Conversely, it is eminently feasible and possible for the radiological and EHR data to be readily accessible. To access data from other healthcare networks, particularly as it relates to identification of IPV, it is likely that legislation will need to be passed, perhaps on a state-by-state basis or mandated in single-payer integrated health networks such as National Health Service. Precedents already exist for the opioid crisis, in which nearly every state in USA has a prescription drug monitoring programme where prescribers can see which patients have obtained opioids from multiple facilities.

Integration of AI

A major challenge, therefore, is how to aggregate all the imaging findings and clinical data in a timely fashion from multiple studies of different body parts at different times and often at different institutions. The radiologists and clinicians often do not have the time or resources to access, review and assimilate all the available information. The imaging and clinical data are currently siloed due to subspecialisation and diverse operating systems. Eighty per cent of IPV patients come through the emergency department, yet only 5%–30% are successfully identified.14 The brief patient interaction constrains the physician to focus only on acute conditions and rely on the information shared by the patient instead of optimising comprehensive care by reviewing historical imaging and clinical data. A 2009 study mining the diagnostic histories, International Classification of Dieases (ICD) diagnostic codes, and the number of annual hospital and emergency departments visits for over 500 000 patients showed that a subset of information commonly held in patients’ electronic medical records can help predict who is at risk of future abuse.15 Given such a conundrum, it would be useful to first define specific evidence-based imaging and clinical correlates of IPV and then generate a clinical decision support rule that will have adequate sensitivity and specificity to identify imaging and clinical findings that are associated with IPV. Even here, defining a gold or reference standard for the purpose of IPV identification is challenging due to the high prevalence of unreported and undocumented cases. A patient in the true negative group could potentially be a false negative as the current system heavily relies on self-reporting and self-disclosure. Just because IPV is not documented in EHR, an injury pattern suspicious of IPV on radiological studies should not be dismissed. It is, therefore, prudent to develop a ‘bronze’ standard serving as a pragmatic starting point by combining the information from the image analysis and clinical chart review to infer the status of IPV and label the cases in order to generate a clinical decision support rule. Once the model is trained and validated on a testing set of cases, a prediction for risk probability can then be calculated for every patient encountering healthcare system at the point of care. These AI-enabled tools can be integrated in a radiology reporting system as well as EHR to make the alert visible to the radiologists and clinicians when they open the patient’s electronic chart. Radiology report templates with prepopulated elements of risk probability of IPV based on the injury pattern seen on prior radiological studies will prompt the radiologist to review the flagged studies and facilitate consistent documentation. A similar approach can be used in EHR by prepopulating risk probability of IPV in the progress note templates derived from the combination of risk factors and prior clinical notes. As newer AI algorithms get better at various tasks, the efforts of the healthcare providers will get augmented with relevant intelligence at the point of care. Likewise, as the database grows in size, the algorithm is expected to learn ‘better’ and accuracy of classification injury is likely to go up. Use of a clinical decision support and rules-based algorithm can be integrated to generate an automated alert to classify injuries for their likelihood of being as a result of IPV with a risk probability of IPV in every patient and alert the clinicians when high-risk patients are identified. The automated alert can be coupled with visualisation of clinical and radiological indicators that will enable the healthcare providers to quickly review the critical risk factors. There have been some successful attempts to develop a multifaceted child abuse clinical decision support system comprised of a trigger system, alerts and a physical abuse order set, resulting in increase in the number of young children identified as having injuries suspicious for physical abuse in general emergency departments.16–19

Conversational guides centred on AI-enabled automated alert

Currently, healthcare providers rely solely on the patient’s wish to share the information. Screening questions might motivate the patient to disclose the information but if the patient decides against disclosing, the provider might not address the obvious but undisclosed concerns about IPV. Healthcare providers may ignore and miss victims of IPV through their own subconscious bias towards the victim’s and abuser’s physical appearance, and educational and socioeconomic background. It is well recognised that providers are usually hesitant to inquire about IPV due to fear of offending patients or their partners. In contrast, automated prediction of IPV based on historical, radiological and clinical data would be unbiased by the interactions with the victim and the abuser. The alert with visualisation of risk factors will provide an objective basis with why the clinicians approach the patients with more confidence in discussing this unspoken issue and breaking the silence that surrounds IPV. This will also profoundly benefit the patients who are unable to share their story due to altered mental status, mental illness or neurological injury.

Designing conversational guides centred around imaging and objective alerts for physicians and social workers to approach the patients who are identified at high risk for IPV but are not forthcoming are necessary to make interventions successful and to educate clinicians regarding IPV. Visually pointing out the injury on imaging studies to the victim may make them more acceptable to disclose their true story. This discussion, often called the ‘caring conversation’, provides an opportunity for the victim to disclose (or not) the facts to the provider who can then act accordingly.


The goal of this viewpoint has been to encourage healthcare providers to recognise the need and opportunity to make the best use of imaging and technology in identifying victims of IPV based on location and pattern of both acute and chronic injuries. While the injury findings currently exist on radiological studies, their significance often goes unnoticed and undervalued. By increasing the awareness of radiologists and other healthcare providers, integrating adult IPV identification into core training, and embracing machine learning tools, healthcare can increase the visibility of this otherwise ‘invisible’ but a highly prevalent critical health issue. Objective information and alert obtained by longitudinal imaging and clinical data can transform the care plan for victims and can also be of help to law enforcement agencies.

Further research is needed to establish the accuracy of imaging patterns in IPV victims and integration of machine learning into the radiology reporting system and clinical workflow.


The authors would like to thank Kathryn M Rexrode, MD (Women's Health, BWH) and Mitchel B Harris, MD (Orthopedics, Mass General Hospital) for their guidance in preparing this manuscript.



  • Twitter @KhuranaBharti

  • Contributors All authors in this paper have appropriately contributed to the manuscript preparation.

  • Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.

  • Competing interests None declared.

  • Patient consent for publication Not required.

  • Provenance and peer review Not commissioned; externally peer reviewed.

  • Data availability statement There are no data in this work.