The burden of trauma has significantly shifted from fatal to non-fatal injuries. In 2015, non-fatal injuries accounted for approximately 31 million injuries of people treated and released from the ER. Although it is difficult to quantify the impact of injury, the use of injury data to define problems and identify intervention strategies may benefit public health. There is a need to capture external causes of non-fatal injuries across two coding systems. Literature does not support a current comparability matrix or tool that builds a crosswalk between National Electronic Injury Surveillance (NEISS) and International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) to determine specific differences in injury diagnosis classification. Unstructured data within medical records create a greater challenge to accurately code injuries based on supplemental information provided within narrative notes.
Project Aim/Purpose:
The aim of this quality improvement project is to implement a comparability tool between NEISS-coded injuries and ICD-10-CM external cause codes (“V”, “W”, “X” and “Y”) that maps non-fatal injuries cases to a three and four character. The purpose of creating a bridging matrix is to standardize data between the two classification systems to allow for more comparable and quantifiable data of non-fatal injuries, as well as establish consistent quality control checks within the National Electronic Injury Surveillance System- All Injury Program Web-based Injury Statistics Query and Reporting System (NEISS-AIP WISQARS) cause coding categories. The project’s goal is to design and implement a bridging matrix between collected National Electronic Injury Surveillance (NEISS) coded-injuries and ICD-10-CM external cause codes that is based on the external cause of injury framework.
Methods:
The design of the text-search algorithm depended heavily on the building of Statistical Analysis System (SAS) programs use of PERL regular expressions to capture semantics and typographical errors found within narratives and other variables to determine external cause of injury. The scope of the project was narrowed to the mechanism-of-injury category, dog bites. The variables included were patient age, gender, affected body part, diagnosis, and discharge disposition. The sample was made up of 613,422 cases non-fatal injuries that were treated in the emergency room and discharged from January 1, 2015 to December 31, 2015. It was established that 5,811 cases represented NEISS injury-coded narratives for external cause code, 16 (dog bites).
Analysis
The number of NEISS-coded injury narratives matched to ICD-10-CM external cause codes served as the numerator. The denominator consisted of the entire list of 2015 NEISS coded cases for dog bite injuries (NEISS code, 16; 5,811 cases). It was important to measure the percentage of matches retrieved between the NEISS-coded injury data and ICD-10-CM comparability tool and the NEISS-AIP WISQARS database to a corresponding mechanism of injury. The number of NEISS-coded injury narratives matched to ICD-10-CM external cause codes represented the numerator. The denominator was the number of dog bite records reported within the WISQARS' 2015 Non-fatal Injury database, capturing only cases where a person was treated and released from the ER.
Results:
Out of the 5811 dog bite narratives, the algorithm successfully identified and associated 5,694 injury narratives (96.8% of all dog bite cases) with ICD-10-CM external cause code, W54.0, using NEISS cause code 16 (dog bite), which contained 3,010 words with dog bite data. There were 424 false positive cases and 117 false negative cases. The precision was 0.93 and the recall was 0.98. The text algorithm also mapped at greater than 92% accuracy to the Web-based Injury Statistics Query and Reporting System (WISQARS). WISQARS determined that out of the 5, 811 potential dog bite cases, 5,445 narrative cases were reportable, non-fatal dog bites in 2015. Patients who sustained a non-fatal dog bite that involved assault, confirmed or suspected and legal intervention were excluded from the final count by WISQARS.
Discussion of Results
The potential of standardizing NEISS-coded injury with ICD-10-CM external cause codes ((“V”, “W”, “X” and “Y”) would allow for new developmental work by practitioners, policy-makers, and researchers in the area of injury prevention. The mapping tool would serve as a new way to perform quality control on NEISS Hospital abstractors for coding accuracy and additional training opportunities. Limitations to the project included missing and inaccurate data identified within NEISS’ variable fields (i.e., patient age, gender, affected body part, diagnosis, discharge disposition, and narrative note). One shortcoming of the NEISS classification system is that only two injuries may be assigned to one case (i.e., cause and cause2 fields). For individuals with multiple injuries, less severe injuries may not be captured by NEISS; therefore, potentially skewing the number of cases within one external cause code category.
Unlike coded data, free text data offers detail and qualitative value that allows the coder to accurately identify the mechanism and intent of injury. The manual review of all data records is not practicable or infallible. Due to the variations in semantics and keystroking (e.g. typographical errors and spacing) found within narratives, a solely automated review process of NEISS-coded narratives would be premature at this time.
Conclusion:
This quality improvement project provided a basic framework for standardizing unstructured data found within ER narrative text. There have been past studies to use text- search algorithms to ascertain injury data from NEISS narratives. This project is one of few to create a mapping tool between NEISS-code injuries and ICD-10-CM external cause codes. The project team has demonstrated that the design and use of PERL regular expression in SAS is a reliable method to capture valuable narrative within unstructured data. In addition to performing manual reviews as needed, the mapping tool provides a more efficient process in performing quality control checks on coded variables (e.g., sex, diagnosis, body part, race, and disposition).