| Abstract |
Background: Perioperative acute kidney injury (AKI) is challenging to predict and a common complication of lower limb arthroplasties. Our aim was to create a machine learning model to predict AKI defined by both serum creatinine (sCr) levels and urine output (UOP) and to investigate which features are important for building the model. The features were divided into preoperative, intraoperative, and postoperative feature sets.
Methods: This retrospective, register-based study assessed 648 patients who underwent primary knee or hip replacement at Oulu University Hospital, Finland, between January 2016 and February 2017. The RUSBoost algorithm was chosen to establish the models, and it was compared to Naïve/Kernel Bayes and support vector machine (SVM). Models of AKI classified by either sCr levels or UOP were established. All the models were trained and validated using a five-fold cross-validation approach. An external test set was not available at the time of this study.
Results: The performance of both the sCr level- and UOP-based AKI models improved when pre-, intra-, and postoperative features were used together. The best sCr level-based AKI model performed as follows: area under receiving operating characteristic (AUROC) of 0.91, (95% CI ± 0.02), area under precision-recall (AUPR) of 0.35 (95% CI ± 0.04) sensitivity of 0.88 (95% CI ± 0.03), specificity of 0.87 (95% CI ± 0.03), and precision o (95% CI ± 0.03). This model correctly classified 22 out of 25 patients with AKI. The best UOP-based AKI model performed as follows: AUROC of 0.98 (95% CI ± 0.02), AUPR of 0.48 (95% CI ± 0.04), sensitivity of 0.88 (95% CI ± 0.02), specificity of 0.93 (95% CI ± 0.03), and precision of 0.34 (95% CI ± 0.04). This model correctly classified 23 out of 26 patients with AKI. In the sCr-AKI models, estimated glomerular filtration rate (eGFR)-related features were most important, and in the UOP-based AKI models, UOP-related features were most important. Other important and recurring features in the models were age, sex, body mass index, ASA status, operation type, preoperative eGFR, and preoperative sCr level. Naïve/Kernel Bayes performed similarly to RUSBoost. SVM performed poorly.
Conclusions: The performance of the models improved after the inclusion of intra- and postoperative features with preoperative features. The results of our study are not generalizable, and additional larger studies are needed. The optimal ML method for this kind of data is still an open research question.
|
