Balancing Baseline Characteristics With SMOTE and Propensity Score Weighting in a Comparative Study of PASCAL Transcatheter Valve Repair and Medical Therapy for Degenerative Mitral Regurgitation
Author(s)
Lisa S. Kemp, PhD1, Sarah Mollenkopf, MPH2, Rebecca Horn, PhD2, Michael Ryan, MS3, William Irish, PhD4.
1Senior Manager, Global Health Economics and Reimbursement, Edwards Lifesciences, Irvine, CA, USA, 2Edwards Lifesciences, Irvine, CA, USA, 3MPR Consulting, Cincinnati, OH, USA, 4East Carolina University, Greenville, NC, USA.
1Senior Manager, Global Health Economics and Reimbursement, Edwards Lifesciences, Irvine, CA, USA, 2Edwards Lifesciences, Irvine, CA, USA, 3MPR Consulting, Cincinnati, OH, USA, 4East Carolina University, Greenville, NC, USA.
Presentation Documents
OBJECTIVES: Managing imbalanced datasets is a critical challenge in non-randomized studies comparing treatments, where under-represented groups can lead to biased conclusions. To prepare for a planned outcomes analysis, we applied the Synthetic Minority Over-sampling Technique (SMOTE) to address significant imbalance between patients with degenerative mitral regurgitation (DMR) treated with the PASCAL transcatheter valve repair system (Edwards Lifesciences, Irvine, CA) in the CLASP IID randomized trial (NCT03706833) and a real-world, medically managed cohort of DMR patients.
METHODS: Data for the PASCAL group were from the CLASP IID trial, while the comparator cohort was derived from the Optum Market Clarity database. After applying physician expert rules and CLASP IID inclusion/exclusion criteria, baseline characteristics still showed significant imbalances between the 149 PASCAL patients and the 1,335 medically managed controls. To address this imbalance, SMOTE was used to oversample the PASCAL group, generating synthetic samples to create a balanced dataset. We then applied propensity score (PS) logistic regression, calculating inverse probability of treatment weights (IPTW) to adjust for confounding. Standardized mean differences (SMD) were used to assess balance, with an SMD < 0.25 indicating acceptable balance.
RESULTS: Before SMOTE application, baseline characteristics between the PASCAL and control groups exhibited substantial imbalance. After applying SMOTE and calculating IPTW, most covariates achieved acceptable balance, with SMDs well within the target range. However, residual confounding persisted for a few variables, highlighting the challenges of eliminating bias in observational studies.
CONCLUSIONS: The application of SMOTE, followed by PS-based weighting, effectively balanced the baseline characteristics between the PASCAL and comparator groups, addressing significant imbalance that could have biased the comparison. Although residual confounding remains a concern, this methodology demonstrates the potential of SMOTE in cohort selection for non-randomized studies. Future research should consider incorporating SMOTE when dealing with imbalanced datasets to improve the robustness of treatment comparisons in real-world data.
METHODS: Data for the PASCAL group were from the CLASP IID trial, while the comparator cohort was derived from the Optum Market Clarity database. After applying physician expert rules and CLASP IID inclusion/exclusion criteria, baseline characteristics still showed significant imbalances between the 149 PASCAL patients and the 1,335 medically managed controls. To address this imbalance, SMOTE was used to oversample the PASCAL group, generating synthetic samples to create a balanced dataset. We then applied propensity score (PS) logistic regression, calculating inverse probability of treatment weights (IPTW) to adjust for confounding. Standardized mean differences (SMD) were used to assess balance, with an SMD < 0.25 indicating acceptable balance.
RESULTS: Before SMOTE application, baseline characteristics between the PASCAL and control groups exhibited substantial imbalance. After applying SMOTE and calculating IPTW, most covariates achieved acceptable balance, with SMDs well within the target range. However, residual confounding persisted for a few variables, highlighting the challenges of eliminating bias in observational studies.
CONCLUSIONS: The application of SMOTE, followed by PS-based weighting, effectively balanced the baseline characteristics between the PASCAL and comparator groups, addressing significant imbalance that could have biased the comparison. Although residual confounding remains a concern, this methodology demonstrates the potential of SMOTE in cohort selection for non-randomized studies. Future research should consider incorporating SMOTE when dealing with imbalanced datasets to improve the robustness of treatment comparisons in real-world data.
Conference/Value in Health Info
2025-05, ISPOR 2025, Montréal, Quebec, CA
Value in Health, Volume 28, Issue S1
Code
MSR50
Topic
Methodological & Statistical Research
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, SDC: Cardiovascular Disorders (including MI, Stroke, Circulatory)