Postmarket surveillance of arthroplasty device components using machine learning methods

Abstract

Purpose: While joint arthroplasty is generally a safe and effective procedure, there are concerns that some devices are at increased risk of failure. Early identification of total hip arthroplasty devices with increased risk of failure can be challenging because devices consist of multiple components, hundreds of distinct components are currently used in surgery, and any estimated effect needs to address confounding due to device and patient factors. The purpose of this study was to assess the effectiveness of machine learning approaches at identifying recalled components listed by the US Food and Drug Administration using data from a US total joint arthroplasty registry.

Methods: An open cohort study was conducted using data (January 1, 2001, to December 31, 2015) from 74 520 implantations and 348 unique components in the Kaiser Permanente Total Joint Replacement Registry. Exposures of interest were device components used in elective primary total hip arthroplasty. The outcome was time to first revision surgery, defined as exchange, removal, or addition of any component. Machine learning methods included regularized/unregularized Cox models and random survival forest.

Results: Among the recalled components detected were ASR acetabular shell/large femoral head, Durom acetabular shell/Metasul large femoral head, and Rejuvenate modular neck stem. The three components not identified were characterized by small numbers of devices recorded in the registry.

Conclusions: The novel approaches to signal detection may improve postmarket surveillance of frequently used arthroplasty devices, which in turn will improve public health.

Keywords: FDA; Postmarket surveillance; arthroplasty; elastic net; machine learning; random forest.