From: Machine learning in perioperative medicine: a systematic review
Author, years | Study design | Objective | Final cohort | Outcomes | Type of ML | Prediction performance | Comparator/control |
---|---|---|---|---|---|---|---|
Lundberg SM, 2019 | Retrospective/observational single center | Development and testing of a ML-based system that predicts the risk of hypoxemia during general anesthesia | 48,069 | Hypoxemia | GB | AUC 0.92 | ML-based system was compared to anesthesiologists’ predictions |
Kendale S, 2018 [13] | Retrospective/observational single center | Prediction of the risk of post-induction hypotension using ML methods | 13,323 | Cardiovascular complications | RF, SVM, GB, BN, LR-EN, regularization, K nearest; linear discrimination analysis; neural nets | AUC GB 0.74 (95% CI, 0.72 to 0.77). RF 0.74 (95% CI, 0.73 to 0.75) | Different ML algorithms were trained to obtain the model with the best performance |
Fernandes MPB, 2021 [14] | Retrospective/observational single center | ML models used to predict postoperative mortality rarely include intraoperative factors. | 5015 | Mortality | logistic regression, RF neural networks, SVM and extreme gradient boosting (XGB). | XGB predicted mortality confidence interval (CI): 0.88 (0.83–0.94) | Different ML algorithms were trained to obtain the model with the best performance |
Cherifa M, 2020 [15] | Retrospective/observational single center | Prediction of acute hypotensive episode | 1151 | Cardiovascular complications | Super Learner (SL) algorithm | SL AUROC 0.890 | Different ML algorithms were trained to obtain the model with the best performance |
Flechet M, 2019 [16] | Prospective/observational single center | Compare diagnostic performances of ML models and physicians in predicting AKI-23 in the 7 days following ICU admission | 252 | Acute kidney injury | ML based AKI predictor | AUROC 0.80 | Physicians’ predictions were compared against the AKI predictor model |
Kang AR, 2020 [17] | Retrospective/observational single center | Prediction of hypotension during anesthesia induction | 222 | Cardiovascular complications | Naïve Bayes, logistic regression, RF, ANN | RF best performance AUC 0.842 | Different ML algorithms were trained to obtain the model with the best performance |
Meiring C, 2018 [18] | Retrospective/observational multicentric | Identification of risk factors for admission in ER/ICU for spine patients | 11150 | ER/ ICU admission | RF, SVM, GB, DECISION TREE, DEEP LEARNING, NNC, Single layer averaged neural network | RF AUC 0.859, NNC AUC0.864; SVM AUC 0.867; adaboost AUC 0.868; deep learning AUC 0.883 | Logistic regression against physiological data alone outperformed APACHE-II (current risk stratification tools) |
Nudel J, 2021 [17] | Retrospective/observational multicentric | Comparison of two ML strategies with conventional statistical models in prediction of surgical complication | 43,6807 | Surgical complications, VTE | GB, ANN | ANN, and XGB, LR achieved similar AUCs 0.65, 0.67 and 0.64 | Different ML algorithms were trained to obtain the model with the best performance |
Lee Hc, 2018 [19] | Retrospective/observational single center | Comparison of ML method with logistic regression analysis to predict AKI after cardiac surgery | 2010 | AKI, mortality | RF, SVM, GB, DECISION TREE, DEEP LEARNING, NNC | Best GB AUC 0.78 | The performance of ML approaches was compared with that of LR analysis |
Bai P, 2020 [20] | Retrospective/observational multicentric | Identification of risk factors of early cerebral infarction and myocardial infarction after CEA with ML method | 443 | Cardiovascular complications | linear SVM, decision tree,RF,ANN, quadratic discriminant analysis, and XGBoost | XGBoost had the highest accuracy | Not applicable |
Tan HS, 2021 [2021] | Retrospective study single center | Use of ML to identify predictive factors for inadequate labor anesthesia | 20,716 | Pain prevention | RF, XGBoost and logistic regression models | All three models performed similarly, with AUC 0.763–0.772 | The performance of ML was compared with regression techniques |
Solomon SC, 2020 [21] | Retrospective and prognostic single center | Prediction of intraoperative bradicardia | 62,182 | Cardiovascular complications | Gradient Boosting Machine (GBM) | AUC of 0.81–0,89 | The performance of ML was compared with regression techniques |
Ko S, 2020 [22] | Retrospective and multicentric | Prediction of postoperative AKI after total knee arthroplasty. | 5757 | AKI | Gradient Boosting Machine (GBM) | AUC of 0,78 | Not applicable |
Lu Y, 2020 [23] | Retrospective single center | Develop ML algorithm for identification of patients requiring admission following elective anterior cruciate ligament (ACL) reconstruction. | 4709 | Length of stay | RF, XGBoost, LDA, AdaBoost | The ensemble model achieved the best AUC 0.76 | Not applicable |
Maheshwari K, 2020 [24] | Observational single center | Using ML to predict intraoperative hypotension | 305 | Cardiovascular complications | Hypotension Prediction Index | 95% confidence interval | Not applicable |
Hill BL, 2019 [25] | Retrospective/observational single center | Develop a model that estimates in-hospital mortality at the end of surgery to quantify the change in risk during the perioperative period. | 53,097 | Mortality | Logistic regression, Elastic Net24 logistic regression, RF, GB. | Best RF 0.932 | Comparison of ML methods with the perioperative score (as ASA physical status score) |
Suhre W,2020 [26] | Retrospective multicentric | Correlation between chronic cannabis use and the risk of postoperative nausea and vomiting (PONV). | 16,245 | PONV | Bayesian additive regression trees (BART) | 90% CI 0.98–1.33 | Not applicable |
Lee HC, 2018 [27] | Retrospective/observational single center | Comparison of ML method with logistic regression analysis to predict AKI after liver transplantation | 1211 | AKI, mortality | RF, SVM, GB, Decision tree, Neural network Classifier, BN, LR-EN, multilayer perceptron | Best GB AUC 0.90 | The performance of ML approaches was compared with that of LR analysis |
Barry GS, 2021 [28] | Retrospective cohort study | Investigate the incidence and factors associated with rebound pain in patients who received a PNB for ambulatory surgery. | 482 | Pain control | Logistic model tree attribute-selected classifier | ROC curve of 0.609 | Not applicable |
Gabriel RA, 2019 [29] | Retrospective/observational single center | Develop a predictive model for determining LOS. | 1018 | LOS | Ridge regression, Lasso, RF | ridge regression 0.761, Lasso 0.752, RF 0.731 | Predictive models using ML techniques were compared to model performances |
Li H, 2020 [30] | Retrospective/observational single center | Development of a predictive model for LOS after total knee arthroplasty | 1826 | LOS | GB | AUC 0.738. | Logistic regression and ML model were compared |
Jungquist CR, 2019 [31] | Retrospective/observational single center | Early detection of respiratory depression using ML models | 60 | Postoperative respiratory complications | SVM | Accuracy of 80% | Not applicable |
Nguyen M, 2020 [32] | Multicentric randomized | Using ML techniques and causal inference methods to detect the dynamic relationship between transfusion ratios and outcomes in trauma patients | 680 | Mortality and hemorrhagic complications | Statistical programming language R | Mortality at AUC 0.89, hemorrhagic complications 1.07 | ML techniques were used to augment the intent-to-treat analysis of PROPPR |
Tourani R,2019 [33] | Retrospective multicentric | In the context of perioperative decision support, understand if the use of intraoperative data improve the performance of 30-day postoperative risk models | 38,045 + 9,044 | Sepsis, septic shock, UTI, PNA, surgical infections | Logistic regression models. | AUC between 0.66 and 0.82 | Not applicable |
Cartailler J,2019 [34] | Clinical trial single center | Use of EEG-patterns to anticipate excessive deep sedation | 80 | Neurological complications | RF | AUC of 0.93 | Not applicable |
Wong WEJ, 2021 [35] | Retrospective/observational single center | Prediction of AKi in ICU | 940 | ICU AKI, hospital and 1 year mortality | Chi-square test, Fisher’s exact test,Mann-Whitney test, independent t test and the Kaplan-Meier curve. | AUROCs of the auxiliary models for ICU AKI were 0.7537, 0.7589, 0.7950, 0.7333 and 0.7654. | Not applicable |
Lee CK, 2021 [36] | Retrospective/observational single center | Prediction of mortality in post-operative patients | 59,985 | Post-operative mortality | Generalized additive models with neural networks (GAM-NNs). | AUC 0.921 | Model performance was compared to a standard LR model |
Jeong YS, 2021 [37] | Retrospective/observational single center | To make a proper model for predicting postoperative major cardiac event (MACE) in ESRD patients undergoing general anesthesia. | 3220 | Cardiovascular complications, mortality | SVM, decision tree, RF, Gaussian naive Bayes (GNB), ANN, LR, XGBoost | RF AUC 0.797 | Different ML algorithms were trained to obtain the model with the best performance |
Filiberto AC, 2021 [38] | Retrospective/observational single center | Postoperative acute kidney injury using ML models | 1531 | AKI | RF | AUC 0.70 | ML models using the perioperative data were compared to models using either preoperative data alone or the ASA physical status classification |
Meyer A, 2018 [39] | Retrospective/observational single center | Use machine learning methods to predict severe complications during and after cardiothoracic surgery. | 11,492 | Postoperative bleeding, AKI, mortality | Deep learning model | AUC 0·09 for bleeding, of 0·18 for mortality, and of 0·25 for AKI | Deep learning methods were compare to established standard-of-care clinical reference tools |
Chiew CJ, 2020 [40] | Retrospective/observational single center | Compare the performance of ML models against the traditionally (CARES) model and (ASA-PS) in the prediction of 30-day postsurgical mortality and ICU admission | 90,785 | Mortality, postoperative ICU admission | RF, GB, adaptive boosting, SVM | Best GB AUC 0.23 and for mortality and 0.38 ICU admission | The performance of ML models was compare against the traditionally Combined Assessment of Risk and Encountered in Surgery (CARES) model and the ASA physical status. |
Bihorac A, 2019 [41] | Retrospective/observational single center | To calculate the risk for postoperative complications and death after surgery using ML | 51,457 | AKI, sepsis, VTE, ICU admission > 48 h, mechanical ventilation > 48 h, wound, neurologic and cardiovascular complications | MySurgeryRisk algorithm | AUC values ranging between 0.82 and 0.94 | Not applicable |
Yao RQ, 2020 [42] | Retrospective/observational single center | Develop a mathematical model for predicting the in-hospital mortality among patients with postoperative sepsis. | 3713 | Postoperative sepsi, mortality | Extreme gradient boosting (XGBoost) and stepwise logistic regression | Best XGBoost AUC 0.835 | ML model was compare to the stepwise LR model. |
Datta S, 2020 [43] | Retrospective/observational single center | Describe a model that predicts postoperative complications considering intraoperative events. | 43,943 | ICU LOS,prolonged mechanical ventilation, neurologic complications cardiovascular complications, AKI, VTE, wound complications, mortality | RF | AUC 0.21 | ML models using preoperative and intraoperative data were compare to models using preoperative data alone |
Brennan M, 2019 [44] | Prospective, non-randomized pilot study | Assess the usability and accuracy of the MySurgeryRisk algorithm for preoperative risk assessment | 20 | AKI, sepsis, VTE, ICU admission > 48 h, mechanical ventilation > 48 h, wound, neurologic and cardiovascular complications | MySurgeryRisk algorithm | MySurgeryRisk algorithm ranged between 0.73 and 0.85 | Compare the accuracy of perioperative risk-assessment between physicians and MySurgeryRisk. |
Houthooft R,2015 [45] | Retrospective/observational single center | develop model to determine patient survival and ICU length of stay (LOS) based on monitored ICU patient data. | 14,480 | LOS | ANN, k-nearest neighbors (k-NN), SVMs, classification trees (CART), RF, adaptive boosting (AdaBoost) | SVM AUC 0.77 | Different ML algorithms were trained to obtain the model with the best performance |