(167) Evaluating Algorithms for Sinus Node Dysfunction in US Health Insurance Claims Databases

Background: Epidemiological safety studies concerning sinus node dysfunction (SND) require accurate computable algorithms, which poses a challenge due to its diverse clinical presentation and variability in electrophysiological findings. No published algorithms reported full performance metrics.

Objectives: To evaluate performance of SND algorithms.

Methods: We systematically built 10 algorithms based on commonly used diagnostic elements (first-time diagnosis of SND, visit context, electrophysiological diagnostic workup, treatment with pacemaker or atropine, and symptoms of fatigue, dizziness, syncope, changes in heart rhythm perception or chest pain). Algorithms varied from a) broad SND diagnosis with hospitalization (common in the literature) to b) narrow SND diagnosis with hospitalization, symptoms within 30 days prior, workup within 30 days prior or after, and treatment within 60 days after the diagnosis. Time windows were chosen based on the 75th percentile of empirical distributions. Across algorithms and in two databases, Merative Medicare (MDCR) and Merative Commercial Claims (CCAE), we systematically characterized patients and estimated performance using OHDSI probabilistic tool PheValuator.

Results: Out of 10 algorithms studied, here we report on the broadest and the narrowest algorithms. The broad algorithm yielded the most patients (153,457 in MDCR and 86,794 in CCAE), while the narrow algorithm (requiring workup, treatment, and symptoms) reduced the number of patients by 63% in MDCR and 82% in CCAE. Treatment was consistently the most limiting criterion (removing 56% and 82% of patients, respectively) and workup was the least (removing 5% of patients). The narrow algorithm resulted in a higher proportion of patients with cardiovascular treatments, diagnoses, and complications then the broad algorithm.
In MDCR and CCAE, the positive predictive value (PPV) of the narrow algorithm was 0.97 (95% Confidence Interval (CI): 0.97 - 0.98) and 0.94 (95% CI: 0.89 - 0.97) respectively though with low sensitivity ( < 0.1). In contrast, the broad algorithm had better sensitivity (0.21 [95% CI: 0.21 - 0.21] and 0.20 [95% CI: 0.19 – 0.20]) and lower PPV (0.93 [95% CI: 0.93 – 093] and 0.67 [95% CI: 0.64 – 0.70], respectively). All algorithms had specificity >0.99.
Across the algorithms, the resulting population differed by database: MDCR captured older patients (mean age 80.0 years vs. 55.4 in CCAE) with more comorbidities and greater healthcare utilization.

Conclusions: We systematically assessed the impact of inclusion criteria on performance of SND algorithms. Such objective examination can guide robust selection of an algorithm that best represents clinical intent and the target population of interest.