Identifying fall-related injuries: Text mining the electronic medical record

Unintentional injury due to falls is a serious and expensive health problem among the elderly. This is especially true in the Veterans Health Administration (VHA) ambulatory care setting, where nearly 40% of the male patients are 65 or older and at risk for falls. Health service researchers and clinicians can utilize VHA administrative data to identify and explore the frequency and nature of fall-related injuries (FRI) to aid in the implementation of clinical and prevention programs. Here we define administrative data as structured (coded) values that are generated as a result clinical services provided to veterans and stored in databases. However, the limitations of administrative data do not always allow for conclusive decision making, especially in areas where coding may be incomplete. This study utilizes data and text mining techniques to investigate if unstructured text-based information included in the electronic medical record can validate and enhance those records in the administrative data that should have been coded as fall-related injuries. The challenges highlighted by this study include data extraction and preparation from administrative sources and the full electronic medical records, de-indentifying the data (to assure HIPAA compliance), conducting chart reviews to construct a ‘gold standard’ dataset, and performing both supervised and unsupervised text mining techniques in comparison with traditional medical chart review.