Comprehensive Summary
In this study, Dragosloveanu et al. evaluated the abilities of eight machine learning (ML) models to classify risk of periprosthetic joint infection (PJI) following joint arthroplasties. Demographics, clinical, and imaging data were collected from 27,854 patients who had undergone joint arthroplasty. Following training and testing, each model was evaluated based on parameters including accuracy, precision, and recall. Findings suggested that Random Forest (RF) performed the strongest, with an accuracy of 0.9975, precision of 0.9957, and recall of 0.9995. RF stood out in its quick computational times, only taking 33.02 seconds for training and 0.53 seconds to generate predictions. While XGBoost and KNN models showed some strengths, RF alone was selected for further trials. Upon clinical employment testing, key predictors used by RF were identified, namely antibiotic count, cardiovascular history, and discharge diagnosis code. ML models have been used previously to identify PJI; however, this study utilized a much larger dataset to expand on previous research.
Outcomes and Implications
PJI is often associated with increased medical costs, higher rates of morbidity, and lower quality of life. The non-specific signs and symptoms of PJI often make it a complication that is hard to identify and especially predict. ML models for PJI risk enables earlier diagnosis and improved patient prognosis. Further studies must be conducted, as this one is limited in its data from one source; however, these findings highlight the potential of RF models to support earlier PJI intervention through risk classification.