LABORATORY-BASED MACHINE LEARNING ALGORITHM REDUCES UNCERTAINTY AND ENHANCES DIAGNOSTIC ACCURACY IN PERIPROSTHETIC JOINT INFECTION
Author(s)
Van Thai-Paquette, MS1, James Parr, PhD2, Amy Worden, BS, MPH3, Krista Toler, MS/MBA4.
1Research Principal Engineer II, Zimmer Biomet, Claymont, DE, USA, 2Machine Learning Lead Engineer, Zimmer Biomet, Swindon, United Kingdom, 3Global Health Economics and Market Access Manager, Zimmer Biomet, Centennial, CO, USA, 4Research & Development Assoc Director, Zimmer Biomet, Warsaw, IN, USA.
1Research Principal Engineer II, Zimmer Biomet, Claymont, DE, USA, 2Machine Learning Lead Engineer, Zimmer Biomet, Swindon, United Kingdom, 3Global Health Economics and Market Access Manager, Zimmer Biomet, Centennial, CO, USA, 4Research & Development Assoc Director, Zimmer Biomet, Warsaw, IN, USA.
OBJECTIVES: Although multiple diagnostic criteria exist to aid physicians in diagnosing periprosthetic joint infection (PJI), gaps in their clinical utility remain. Incorrect implementation by physicians, combined with all leading criteria placing a high number of patients into an “inconclusive” class, collectively hinders timely and confident clinical decision-making. This study aims to evaluate whether a clinical laboratory-based machine learning (ML) model (“Model”) that generates a PJI probability score from synovial fluid (SF) biomarkers can mitigate these implementation gaps.
METHODS: 274 clinical vignettes were previously presented to twelve physicians with varying expertise and experience to assess their ability to diagnose PJI when all necessary information is available. For this study, the same vignettes were evaluated using the Model, and diagnostic performance was compared to the physicians. To generate the PJI probability score, any missing SF biomarker was imputed using a standard method, and SF C-reactive protein (CRP) was converted from serum CRP using a published regression equation. Diagnostic performance was assessed using the 2018 International Consensus Meeting (ICM) Criteria-based definition of PJI, which is the current standard-of-care (SOC).
RESULTS: Overall physician uncertainty rate was 23%, compared to 0.4% by the Model. Physician undecided rate for the ICM Inconclusive patients (n = 30) was very high (45%), compared to just one (3%) equivocal result by the Model. When required to classify all patients, agreement rates to ICM were: Model - 97%; Physicians - 92% (p < 0.001). Nearly all (92%) incorrect classifications by physicians were false positives.
CONCLUSIONS: The ML Model reduced diagnostic uncertainty by >50-fold, rendering only one equivocal result. Model agreement with the current SOC PJI definition was significantly higher than physicians, confirming the implementation gap. Based on these results, implementing this ML Model into clinical practice would prevent over-diagnosis of PJI, reducing unnecessary 2-stage revision surgery and over-prescription of high-dose systemic antibiotics.
METHODS: 274 clinical vignettes were previously presented to twelve physicians with varying expertise and experience to assess their ability to diagnose PJI when all necessary information is available. For this study, the same vignettes were evaluated using the Model, and diagnostic performance was compared to the physicians. To generate the PJI probability score, any missing SF biomarker was imputed using a standard method, and SF C-reactive protein (CRP) was converted from serum CRP using a published regression equation. Diagnostic performance was assessed using the 2018 International Consensus Meeting (ICM) Criteria-based definition of PJI, which is the current standard-of-care (SOC).
RESULTS: Overall physician uncertainty rate was 23%, compared to 0.4% by the Model. Physician undecided rate for the ICM Inconclusive patients (n = 30) was very high (45%), compared to just one (3%) equivocal result by the Model. When required to classify all patients, agreement rates to ICM were: Model - 97%; Physicians - 92% (p < 0.001). Nearly all (92%) incorrect classifications by physicians were false positives.
CONCLUSIONS: The ML Model reduced diagnostic uncertainty by >50-fold, rendering only one equivocal result. Model agreement with the current SOC PJI definition was significantly higher than physicians, confirming the implementation gap. Based on these results, implementing this ML Model into clinical practice would prevent over-diagnosis of PJI, reducing unnecessary 2-stage revision surgery and over-prescription of high-dose systemic antibiotics.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
CO173
Topic
Clinical Outcomes
Topic Subcategory
Comparative Effectiveness or Efficacy, Performance-based Outcomes
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, SDC: Infectious Disease (non-vaccine), SDC: Injury & Trauma