Abstract
Introduction
Next generation sequencing (NGS) has been shown to facilitate detection of microbes in a clinical sample, particularly in the setting of culture-negative periprosthetic joint infection (PJI). However, it is unknown whether every microbial DNA signal detected by NGS is clinically relevant. This multi-institutional study was conceived to 1) identify species detected by NGS that may predict PJI, then 2) build a predictive model for PJI in a developmental cohort; and 3) validate the predictive utility of the model in a separate multi-institutional cohort.
Methods
This multicenter investigation involving 15 academic institutions prospectively collected samples from 194 revision total knee arthroplasties (TKA) and 184 revision hip arthroplasties (THA) between 2017–2019. Patients undergoing reimplantation or spacer exchange procedures were excluded. Synovial fluid, deep tissue and swabs were obtained at the time of surgery and shipped to MicrogenDx (Lubbock, TX) for NGS analysis. Deep tissue specimens were also sent to the institutional labs for culture. All patients were classified per the 2018 Consensus definition of PJI. Microbial DNA analysis of community similarities (ANCOM) was used to identify 17 candidate bacterial species out of 294 (W-value >50) for differentiating infected vs. noninfected cases. Logistic Regression with LASSO model selection and random forest algorithms were then used to build a model for predicting PJI. For this analysis, ICM classification was the response variable (gold standard) and the species identified through ANCOM were the predictor variables. Recruited cases were randomly split in half, with one half designated as the training set, and the other half as the validation set. Using the training set, a model for PJI diagnosis was generated. The optimal resulting model was then tested for prediction ability with the validation set. The entire model-building procedure and validation was iterated 1000 times. From the model set, distributions of overall assignment rate, specificity, sensitivity, positive predictive value (PPV) and negative predicative value (NPV) were assessed.
Results
The overall predictive accuracy achieved in the model was 75.9% (Figure 1). There was a high accuracy in true-negative and false-negative classification of patients using this predictive model (Figure 2), which has previously been a criticism of NGS interpretation and reporting. Specificity was 97.1%, PPV was 75.0%, and NPV was 76.2%. On comparison of the distribution of abundances between ICM-positive and ICM-negative patients, Staphylococcus aureus was the strongest contributor (F=0.99) to the predictive power of the model (Figure 3). In contrast, Cutibacterium acnes was less predictive (F=0.309) and noted to be abundant across both infected and noninfected revision TJA samples.
Discussion
This study is the first to utilize predictive modeling algorithms on a large prospective multicenter database in order to transform analytic NGS data into a clinically relevant diagnostic signal. Our collaborative findings suggest the microbial DNA signal identified on NGS may be an independent useful adjunct for the diagnosis of PJI, as well as help identify causative organisms. Further work applying artificial intelligence tools will improve accuracy, predictive power and clinical utility of high-throughput sequencing technology.
For figures, tables, or references, please contact authors directly.
