머신러닝을 이용한 패혈증 진단 코딩 오류 탐지 모형 개발 (Development of a Model for Detecting Coding Errors in Sepsis Diagnosis Using Machine Learning)

서지정보

· 발행기관 : 한국병원경영학회
· 수록지 정보 : 병원경영학회지 / 30권 / 2호 / 50 ~ 60페이지
· 저자명 : 신동교, 김지만

초록

Purposes: This study aims to develop a machine learning-based model to detect diagnosis coding errors in the diagnosis-related group (DRG)-based payment system in Korea, focusing on sepsis—a diagnosis with the highest adjustment rate.
Methodology/Approach: We analyzed five years of inpatient claims data (2015–2019) from a tertiary hospital. A total of 1,068 claims involving sepsis as either a principal or secondary diagnosis were examined using the density-based spatial clustering of applications with noise (DBSCAN) algorithm. Abnormal cases were identified based on discrepancies between DRG-based payments and adjusted fee-for-service (FFS) charges. Key covariates such as age, comorbidity index, length of stay, department, physician, and coder characteristics were controlled for using linear regression. Model performance was evaluated against Health Insurance Review and Assessment Service (HIRA) audit results.
Findings: For sepsis coded as a principal diagnosis, the model achieved a balanced accuracy of 55.0% and an overall accuracy of 70.8%. For secondary diagnoses, the model showed a balanced accuracy of 98.9% and an overall accuracy of 97.8%. Notably, the lift value for detecting errors in secondary diagnoses was 22.73 times higher than that of conventional random audits, indicating substantial improvement in screening efficiency under limited review resources.
Practical Implications: The proposed model may serve as a first-tier screening tool for HIRA audits, prioritizing high-risk claims. It can also support coder training and discharge documentation review. By integrating machine learning into the DRG monitoring process, this approach offers a scalable and data-driven strategy to improve coding accuracy and ensure sustainable reimbursement operations.

영어초록

Purposes: This study aims to develop a machine learning-based model to detect diagnosis coding errors in the diagnosis-related group (DRG)-based payment system in Korea, focusing on sepsis—a diagnosis with the highest adjustment rate.
Methodology/Approach: We analyzed five years of inpatient claims data (2015–2019) from a tertiary hospital. A total of 1,068 claims involving sepsis as either a principal or secondary diagnosis were examined using the density-based spatial clustering of applications with noise (DBSCAN) algorithm. Abnormal cases were identified based on discrepancies between DRG-based payments and adjusted fee-for-service (FFS) charges. Key covariates such as age, comorbidity index, length of stay, department, physician, and coder characteristics were controlled for using linear regression. Model performance was evaluated against Health Insurance Review and Assessment Service (HIRA) audit results.
Findings: For sepsis coded as a principal diagnosis, the model achieved a balanced accuracy of 55.0% and an overall accuracy of 70.8%. For secondary diagnoses, the model showed a balanced accuracy of 98.9% and an overall accuracy of 97.8%. Notably, the lift value for detecting errors in secondary diagnoses was 22.73 times higher than that of conventional random audits, indicating substantial improvement in screening efficiency under limited review resources.
Practical Implications: The proposed model may serve as a first-tier screening tool for HIRA audits, prioritizing high-risk claims. It can also support coder training and discharge documentation review. By integrating machine learning into the DRG monitoring process, this approach offers a scalable and data-driven strategy to improve coding accuracy and ensure sustainable reimbursement operations.

참고자료

· 없음