Improving the accuracy of stroke clinical coding with open-source software and natural language processing

Abstract Clinical coding is an important task, which is required for accurate activity-based funding. Natural language processing may be able to assist with improving the efficiency and accuracy of clinical coding. The aims of this study were to explore the feasibility of using natural language processing for stroke hospital admissions, employed with open-source software libraries, to aid in the identification of potentially misclassified (1) category of Adjacent Diagnosis Related Groups (ADRG), (2) the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision, Australian Modification (ICD-10-AM) diagnoses, and (3) Diagnosis Related Groups (DRG). Data was collected for consecutive individuals admitted to the Royal Adelaide Hospital Stroke Unit over a five-month period for misclassification identification analysis. 152 admissions were included in the study. Using free-text discharge summaries, a random forest classifier correctly identified two cases classified as B70 (“Stroke and Other Cerebrovascular Disorders”) that should be classified as B02 (having received endovascular thrombectomy). A regular expression-based analysis correctly identified 33 cases in which ataxia was present but was not coded. Two cases were identified that should have been classified as B70D, rather than B70A/B/C, based on transfer to another centre within five days of admission. A variety of techniques may be useful to help identify misclassifications in ADRG, ICD-10-AM and DRG codes. Such techniques can be implemented with open-source software libraries, and may have significant financial implications. Future studies may seek to apply open-source software libraries to the identification of misclassifications of all ICD-10-AM diagnoses in stroke patients.