Monitoring of blood biochemical markers for periprosthetic joint infection using ensemble machine learning and UMAP embedding.

introductionPeriprosthetic joint infection (PJI) is a serious complication after total joint arthroplasty. It is important to accurately identify PJI and monitor postoperative blood biochemical marker changes for the appropriate treatment strategy. In this study, we aimed to monitor the postoperative blood biochemical characteristics of PJI by contrasting with non-PJI joint replacement cases to understand how the characteristics change postoperatively. MATERIALS AND

methodsA total of 144 cases (52 of PJI and 92 of non-PJI) were reviewed retrospectively and split into development and validation cohorts. After exclusion of 11 cases, a total of 133 (PJI: 50, non-PJI: 83) cases were enrolled finally. An RF classifier was developed to discriminate between PJI and non-PJI cases based on 18 preoperative blood biochemical tests. We evaluated the similarity/dissimilarity between cases based on the RF model and embedded the cases in a two-dimensional space by Uniform Manifold Approximation and Projection (UMAP). The RF model developed based on preoperative data was also applied to the same 18 blood biochemical tests at 3, 6, and 12 months after surgery to analyze postoperative pathological changes in PJI and non-PJI. A Markov chain model was applied to calculate the transition probabilities between the two clusters after surgery.

resultsPJI and non-PJI were discriminated with the RF classifier with the area under the receiver operating characteristic curve of 0.778. C-reactive protein, total protein, and blood urea nitrogen were identified as the important factors that discriminates between PJI and non-PJI patients. Two clusters corresponding to the high- and low-risk populations of PJI were identified in the UMAP embedding. The high-risk cluster, which included a high proportion of PJI patients, was characterized by higher CRP and lower hemoglobin. The frequency of postoperative recurrence to the high-risk cluster was higher in PJI than in non-PJI.

conclusionsAlthough there was overlap between PJI and non-PJI, we were able to identify subgroups of PJI in the UMAP embedding. The machine-learning-based analytical approach is promising in consecutive monitoring of diseases such as PJI with a low incidence and long-term course.