Limitations of current risk-adjustment models in the era of coronary stenting

Abstract

Background: Several risk-adjustment models have been developed to compare outcomes of conventional coronary angioplasty across physicians and institutions. Yet the accuracy of these models in contemporary interventional practice--characterized by the widespread use of stents and novel adjuvant pharmacotherapies--has not been sufficiently studied.

Methods: The principal published predictive models for inhospital mortality after angioplasty were validated in 11,681 patients undergoing coronary stenting and 6475 patients undergoing balloon-only procedures in the Society for Cardiac Angiography and Interventions registry from July 1996 to December 1998. We examined the 2 components of model accuracy: discrimination, as determined by the c-index; and calibration, as measured by the Hosmer-Lemeshow statistic and predicted-versus-observed probability plots.

Results: The discriminative properties of the models were preserved in the validation cohort and did not differ statistically from one another (c-indexes 0.85-0.89). Hosmer-Lemeshow statistics, however, showed poor fit (P <.001), with all 3 models substantially overestimating the risk of adverse outcomes. Although recalibration of the models achieved satisfactory goodness of fit, laboratory-specific ratings differed depending on the model applied.

Conclusions: Predictive models developed in the era of conventional angioplasty cannot be applied directly to current interventional practice. Although recalibration restores model fit, application of different recalibrated models yields inconsistent assessment of laboratory performance. Development of new, widely generalizable models is warranted, but such models will require continued reassessment as medical technology evolves and practice patterns change.