Creation and Validation of an Algorithm for Predicting the Recurrence of Atrial Fibrillation Following Pulmonary Vein Isolation by Utilizing Real-World Data and Ensemble Modeling Techniques

Introduction Catheter ablation (CA) of atrial fibrillation (AF) represents a mainstay in the treatment of this increasingly prevalent arrhythmia. Prospective clinical trials investigating the efficacy of CA may poorly represent real-world patient populations. However, many real-world clinical datasets possess missing data, which may impede their applicability in research. Thus, we sought to use ensemble modeling to address missing data and develop a model to estimate the probability of AF recurrence following CA. Methods We retrospectively analyzed clinical variables in 476 patients who underwent an initial CA of AF. Univariate and multivariate logistic regression was performed to determine those variables predictive of AF recurrence. A multivariate logistic model was created to estimate the probability of AF recurrence after CA. Missing data were addressed using ensemble modeling, and variable selection was performed using the aggregate of multiple models. Results After analysis, six variables remained in the model: AF during the post-procedural blanking period, coexistence of atrial flutter, end-stage renal disease, reduced left ventricular ejection fraction, prior failure of anti-arrhythmic drugs, and valvular heart disease. Predictive modeling was performed using these variables for 1000 randomly partitioned datasets (80% training, 20% testing) and 1000 random imputations for each partitioned dataset. The model predicted AF recurrence with an accuracy of 74.34 ± 3.99% (recall: 54.03 ± 8.15%; precision: 89.30 ± 4.21%; F1 score: 81.08 ± 3.65%).  Conclusion We successfully identified six clinical variables that, when modeled, predicted AF recurrence following CA with a high degree of classification accuracy. Application of this model to patients undergoing CA of AF may help identify those at risk of post-procedural AF recurrence.

Measuring standards of care for early breast cancer in an insured population

PURPOSE

To demonstrate the use of a combined data base to evaluate the care for local/regional invasive breast cancer in a large insured population of women aged less than 64 years.

PATIENTS AND METHODS

We linked the procedural and hospital claims from Blue Cross Blue Shield (BCBS) of Virginia with clinical stage data from the Virginia Cancer Registry (VCR) from 1989 to 1991. A total of 918 women were assessed with a median age of 50 years; 68% had tumors less than 2 cm, 30% had positive axillary nodes, and 68% were assessed as having local summary stage. A quality-of-care "report card" was used based on standards of care from international Consensus Conferences.

RESULTS

Eight percent had a mastectomy as the initial biopsy procedure. Sixty-nine percent of women ultimately underwent mastectomy. Of those women who underwent lumpectomy, 86% had subsequent radiation. Within 3 months of diagnosis, 43% had a bone scan and 20% a computed tomography (CT) scan. Of women with positive axillary lymph nodes, 83% aged less than 51 years and 52% aged 51 to 64 years received chemotherapy. Fifty-six percent of all women had claims from a medical oncologist. Of women having a total mastectomy, 27% had claims from a plastic surgeon. Sixty-six percent to 76% of women had a mammogram, 24% a bone scan, and 14% a CT scan in the 0-18 and 18-36 month intervals following primary treatment.

CONCLUSION

This study confirms the feasibility of linking sources of data that provide complementary information needed to develop measurements regarding standards of quality and efficiency of oncologic care. This report should serve as an initial benchmark while we await reports from other populations to define the best practice.