Introduction

Violent injury is a leading cause of morbidity and mortality in the USA.1 Among African American adolescents, homicide is the leading cause of death1 ; and for every homicide there are 94 non-fatal violent injuries.2 A majority of the violent injuries treated in the emergency department (ED) are not reported to law enforcement.3 4

Injury surveillance describes the relative magnitude of injury type, monitors trends in injuries, identifies new injury burdens and evaluates prevention and intervention efforts; injury surveillance data are subsequently used to inform research and intervention.5 Healthcare providers are required to use the International Statistical Classification of Diseases and Related Health Problems, 10th Revision, Clinical Modification (ICD-10-CM) to report medical data to the US Department of Health and Human Services.6

A recent study evaluating the impact of the transition from ICD-9-CM to ICD-10-CM confirmed the increased specificity of ICD-10-CM by decreasing the use of codes with undetermined intent and increased identification of assaults based on injury codes.7 The degree to which these ICD-10-CM external cause codes are reliably used in clinical practice to identify intentional injury is unclear.

This study provides insight into a population identified through a violence prevention effort that specifically and prospectively identifies victims of assault or interpersonal violence, rather than relying on external cause codes to identify mechanism and intent. While one may infer a stab wound or gunshot wound results from violence, a diagnosis of ‘head injury’ could result from a motor vehicle crash, a sports injury or a physical assault. This study analyses ICD-10-CM injury codes in patients with a self-reported injury screener to determine the effectiveness of provider-assigned ICD-10-CM coding in capturing injury intentionality.

Methods

Patients presenting to the ED of a large, Southeastern US urban hospital were screened for intentional injuries. The ED is located at a designated level I trauma centre with an annual census of over 150 000 patient visits per year. Approximately 70%–86% of ED patients were screened by a nurse verbally asking the patient questions relating to injury and interpersonal violence as part of the Cardiff Violence Prevention Model, a violence prevention programme.8 Answers were recorded in the electronic medical record.

The dataset consisted of 380 922 ED records for adult patients who were screened at triage between 20 November 2015 and 30 September 2019. Data included demographics (age, race and sex), responses to the injury screener and provider-assigned ICD-10-CM codes.8 Providers, including residents, nurse practitioners, physician assistants and attending physicians, assigned final diagnosis ICD-10-CM code(s) for each patient at the end of the patient’s encounter.

The injury screen contains two questions analysed here: ‘were you injured?’ and ‘was someone trying to hurt you?’ Records were identified as containing a self-reported injury if ‘yes’ was answered for the first question and as assault or interpersonal violence cases if they answered ‘yes’ to both. To identify cases coded for injury and assault using ICD-10-CM diagnostic codes, every ICD-10-CM code assigned to each record was scanned for any relevant injury diagnosis codes (table 1). Records were identified as an assault if they contained any assault codes from the ICD-10-CM.9 Using self-reported injury and assault as reference standards, sensitivity and specificity were calculated for ICD-10-CM codes capturing injury and assault respectively in the provider-coded electronic health records. Data analysis was conducted using SAS V.9.4.

Neither patients or the public were involved in the design, conduct, reporting or dissemination plans of our research.

Result

In the 380 922 records examined, 32 788 (8.61%) patients self-reported any injury and a subset of 6763 (1.78%) reported assault. Additionally, by ICD-10-CM code identification, 57 656 (15.14%) injury cases and 258 (0.07%) assault cases were identified. Patients were predominantly non-Hispanic black (76.9%; n=2 92 875) and male (57%; n=2 17 905); see table 2.

Table 3 contains the sensitivity, specificity and predictive values of both injury and assault in ICD-10-CM codes. For any injury, ICD-10-CM codes had a sensitivity of 67.40% (95% CI 66.89% to 67.91%) and specificity of 89.79% (95% CI 89.69% to 89.89%). Positive predictive value was 38.3% (95% CI 37.93% to 38.73%) and negative predictive value was 96.69% (95% CI 96.63% to 96.75%). For assault, ICD-10-CM codes had a sensitivity of 2.25% (95% CI 1.91% to 2.63%]) and specificity of 99.97% (95% CI 99.97% to 99.98%]). Positive predictive value was 58.91% (95% CI 52.64% to 64.98%) and negative predictive value was 98.26% (95% CI 98.22% to 98.3%).

Both self-report (screener) and ICD-10-CM codes identified injury and violence, and 22 099 records were identified by both methods, 35 557 were identified via ICD-10-CM code alone and 10 689 were identified by injury screener alone. There were distinct instances where an injury was only found via one method. The top 10 primary diagnosis codes were reviewed (table 4). The top primary diagnoses in cases identified only by ICD-10-CM code were S-codes (eg, injury) with many related to injuries to the head (eg, open wound of head, superficial injury of head), whereas the top diagnoses for cases identified via the injury screener alone contained a variety of codes from other ICD-10-CM chapters largely related to pain in a region of the body instead of an injury.

Discussion

Provider-assigned ICD-10-CM codes did not reliably identify injuries and assaults that were self-reported in an ED screen. Although specificity was high for ICD-10-CM codes to identify injury, the sensitivity was less than 70% despite injury codes identifying almost twice the number of encounters for injury as those actually self-reported (15.14% and 8.61%, respectively). Based on this analysis, both ICD-10-CM codes and self-report have overlap and complement each other with additional identification of injury. For surveillance purposes, this is a concern because certain types of injuries are then likely to be systematically overlooked and trends that may reflect risks to public health may not be detectable if relying exclusively on one method.

Assault was poorly identified by provider-assigned ICD-10-CM external cause codes, which had only a 2.25% sensitivity for self-reported assault injury. This demonstrates a severe barrier to identification and monitoring of the violent injury burden and raises concern for underestimation of incidence and prevalence. Intentionality components of injury codes may be even more frequently included in this analysis than in typical practice of injury surveillance by public health entities as it included not only the primary diagnosis code, but all ICD-10-CM codes associated with the encounter. Future research may consider examining the differences in sensitivity when performed by professional coders as compared with provider coded records.

This study has several limitations as this was performed at a single, urban, level 1 trauma-designated centre and may not reflect the broader injury and assault trends across the nation. Also, providers assigned diagnostic codes rather than using medical coding services. We expect this may limit the number of diagnoses applied given only one code is required to complete the patient encounter and additional time is required to record qualifiers associated with a diagnosis. The implications would be better understood with broader study of how ICD-10-CM codes are assigned beyond this single centre. Nurses are trained to request violent injury information and patients only reveal information they feel comfortable sharing. Increased nurse refresher training on the screen may be needed to improve mechanism reporting. More exploration of patient perceptions of the screen and hesitancy to report intentionality or mechanism could be a focus of future research.

Despite these limitations, these data show high specificity and negative predictive values for injury and assault. More work is needed to understand what factors most affect sensitivity such as who assigns codes (provider or medical coding service), coding practices and training for ICD-10-CM. If high specificity can be maintained and sensitivity significantly increased, ICD-10-CM codes may provide important surveillance information for injury and assault patterns.

Conclusion

In this study, provider-assigned ICD-10-CM codes and self-reported injury both demonstrated gaps in surveillance. Improvements are needed in implementation of ICD-10-CM use, especially to improve their reliability for assault injuries. Injury surveillance methods should seek to improve ICD-10-CM use and explore additional measures for the identification and trending of violence and injuries for population health.