Table 5 A summary on the details of ML, strengths and limitations of the studies on diagnosis of PJI
First author [Ref.] | ML application | Algorithm | Statistical performance | Strengths | Limitations |
|---|---|---|---|---|---|
Kuo, F.C. (2021) [8] | A personalized PJI diagnosis | Two-level stacked generalization architecture: -Meta-classifier: Support vector machine; -Base classifiers: Randomforest, eXtreme gradient boosting, logistic regression, naïve bayesian | AUC: 0.988. Accuracy: 96.4%. Recall: 0.981. F1 score: 0.97. Matthews correlation -coefficient: 0.926. Precision: 0.96 | The performance outperformed that of International Consensus Meeting criteria; if-then rule was used for the explanation of the results | No external validation; small cohort size |
Tao, Y. (2022) [13] | PJI pathological diagnosis | ResNet34 deep learning convolutional network | AUC: 0.8136. Average accuracy: 93.3%. Average recall rate: 0.9739. F1 score: 0.9482 | External validation was conducted | Small cohort size |