Back to 2024 Display Posters
Optimizing Claims Codes To Identify Claudication And Chronic Limb-threatening Ischemia Using A Machine-learning Approach
Sanuja Bose, MD, MPH1, Sharon C. Kiang, MD
2, Daniel Roh
2, Jialin Mao, MD, PhD
3, Andrew Cabrera
2, David P. Stonko, MD, MS
1, James H. Black, III, MD
1, Jesse A. Columbo, MD
4, Philip P. Goodney, MD
4, Leigh Ann O'Banion, MD
5, Roger T. Tomihama, MD
2, Caitlin W. Hicks, MD, MS
1.
1The Johns Hopkins University School of Medicine, Baltimore, MD, USA,
2Loma Linda University Medical Center, Loma Linda, CA, USA,
3Weill Cornell Medical College, New York, NY, USA,
4Dartmouth-Hitchcock Medical Center, Lebanon, NH, USA,
5University of California San Francisco - Fresno, Fresno, CA, USA.
OBJECTIVES: The accuracy of using contemporary claims codes to identify patients with peripheral artery disease is controversial. We aimed to validate a predefined set of ICD-10 codes used to identify patients with claudication and chronic limb-threatening ischemia (CLTI), and to optimize their diagnostic accuracy using a supervised machine-learning (ML) approach.
METHODS: We included all patients who underwent a peripheral vascular intervention for claudication or CLTI in the Vascular Quality Initiative (VQI) Implant Surveillance and Interventional Outcomes Network database between January 2016-December 2019. Gold standard claudication and CLTI diagnoses were determined using VQI registry data and compared to a predetermined set of ICD-10 codes for all patients in the Medicare-matched cohort. We used traditional logistic regression models and six ML models (random forest, gradient boosting, decision tree, gaussian-naïve Bayes, multi-layer perceptron, and logistic regression classifier) to distinguish claudication from CLTI using an expanded set of ICD-10 codes. We evaluated sensitivity, specificity, accuracy, and AUC for all models, implementing grid search cross-validation to boost the performance of the ML models.
RESULTS: Of 50,003 patients who underwent a peripheral vascular intervention, 20,769 (38.3%) had claudication and 33,411 (61.7%) had CLTI per gold standard VQI definitions. The predefined set of ICD-10 codes had high sensitivity (80.9%), specificity (81.9%), accuracy (81.3%), and AUC (0.87) for distinguishing claudication versus CLTI. Traditional logistic regression improved sensitivity to 96.2%, but with a substantial drop in specificity (41.8%). Of the ML models, gradient booster classifier performed the best with sensitivity improved to 88.6%, accuracy to 84.2%, and AUC to 89.2%, with minimal drop in specificity (77.1%) (
Table). Permutation feature importance identified 318 ICD-10 codes that significantly improved the predictive algorithm.
CONCLUSIONS: ICD-10 claims codes can be successfully used to accurately discriminate claudication from CLTI without the need for granular registry follow up data. Different prediction models can be used to optimize accuracy, sensitivity, or specificity depending on the study design.
Back to 2024 Display Posters
