Analytical agreement and decision-threshold concordance between Abbott ARCHITECT and Thermo Fisher KRYPTOR procalcitonin assays

Abstract Objectives Procalcitonin (PCT) is used as an adjunct in infection assessment, sepsis risk stratification, and antibiotic stewardship. This study evaluated numeric agreement and threshold-level concordance of the Abbott ARCHITECT and Thermo Fisher KRYPTOR PCT assays at commonly used decision thresholds. Methods A cross-sectional paired method-comparison study was performed using a convenience sample of 130 unique eligible residual clinical serum specimens, one per patient, selected from specimens available during routine method-verification activities. Each specimen was tested from the same thawed aliquot on both platforms. Agreement was assessed using Passing–Bablok regression with 5000 paired-bootstrap resamples, concentration-dependent regression-based agreement bounds, and binary concordance at 0.5, 2.0, and 10.0 ng/mL. In an exploratory post hoc sensitivity analysis, observations with an absolute standardized residual >3 were flagged without reference to decision-threshold classification; all observations remained in the primary analysis. Clinical trial registration this retrospective residual-specimen laboratory method-comparison study was not a prospective clinical trial and was not registered in a trial registry. Results Passing–Bablok regression yielded ARCHITECT = 1.051 × KRYPTOR−0.002; the slope 95% CI was 1.040–1.070 and the intercept 95% CI was − 0.0028 to 0.0002 ng/mL. The expected ARCHITECT-minus-KRYPTOR difference increased from − 0.037 ng/mL at a paired mean of 0.5 ng/mL to 0.053 ng/mL at 2.0 ng/mL and 0.531 ng/mL at 10.0 ng/mL, with concentration-dependent 95% agreement bounds. Kappa values were 0.953, 0.909, and 1.000 at 0.5, 2.0, and 10.0 ng/mL, with 3, 5, and 0 discordant classifications, respectively. Among 30 KRYPTOR-high specimens at 10.0 ng/mL, the exact 95% CI for the unobserved discordance rate was 0–11.6%. Conclusions Threshold classifications were closely aligned in this dataset, but proportional bias and widening concentration-dependent differences preclude a claim of unrestricted numerical interchangeability. Serial monitoring should preferably remain on the same platform, and uncertainty at the high cutoff should not be inferred from zero observed discordance alone.