Automated ICD-9 and ICD-10 Coding with Machine Learning: A Real-World Study Using Electronic Medical Record Text from Udon Thani Cancer Hospital, Thailand

Manually identifying International Classification of Diseases (ICD) codes-of which there are more than $\mathbf{7 0, 0 0 0}$ codefrom text of medical records is both challenging and timeconsuming. Each medical record can require 5-10 minutes to identify, leading to bottlenecks in many hospitals. This study aimed to develop predictive models for automated ICD-9 and ICD10 coding. We conducted eight experiments covering six key medical code categories: ICD-10 Principal Diagnosis, Comorbidity Diagnosis, Complication Diagnosis, External Cause Diagnosis, Other Diagnosis, and ICD-9 Procedure Diagnosis. Depending on the labeling scheme, the models were formulated as either singlelabel or multi-label classification tasks. Our results show that three machine learning models-Random Forest, Support Vector Machine, and Multilayer Perceptron-demonstrate promising performance. The highest performance, with an F 1-score of 0.95, was achieved by the multi-label classification (255 unique classes) for the “Other Diseases” group in ICD-9 Procedure Diagnosis Coding. In contrast, the lowest performance, with an F1-score of 0.53, was observed in the multi-label classification (483 unique classes) for the “Other Diseases” group in ICD-10 Comorbidity Diagnosis Coding.