Table 3.

 A summary of reviewed studies on predicting postoperative outcomes of total knee arthroplasty

Author (Year)	Journal	Prediction outcome	AI/ML algorithm(s)	Statistical performance	Strengths	Weaknesses	Clinical significance of study
Huber (2019) [28]	BMC Medical Informatics and Decision Making	Postoperative improvement in PROMs	Extreme gradient boosting, multi-step adaptive elastic-net, random forest, neural net, Naïve Bayes, k-Nearest Neighbors	AUCs: 0.86 (VAS) & 0.70 (Q score).	Comparison of a wide variety of ML approaches in addition to regression methods	Training and testing sets were selected from the same dataset.	Identified important predictors for postoperative PROMs (e.g., preoperative VAS).
Harris (2021) [20]	The Journal of Arthroplasty	Postoperative 1-year achievement of MCID	LASSO regression, GBM, quadratic discriminant analysis	AUC: 0.76 (ADL), 0.72 (pain), 0.72 (symptoms), 0.71 (quality of life).	Provided sensitivity and specificity of various thresholds of predicted probability of failure to achieve MCID 1 year post-TKA.	Training and testing sets were selected from the same dataset.	Demonstrated potential for AI to predict patients most likely to benefit from TKA.
Kunze (2020) [25]	The Journal of Arthroplasty	Postoperative patient dissatisfaction	Stochastic gradient boosting, random forest, support vector machine, neural network, elastic-net penalized logistic regression	AUC: 0.66–0.79.	All five machine learning algorithms demonstrated superior predictive performance than the standard logistic regression model. Algorithm-identified predictors of postoperative patient dissatisfaction are consistent with previous systematic reviews.	Training and testing sets were selected from the same dataset.	Demonstrated potential for AI to predict patients most likely to experience postoperative dissatisfaction.
Farooq (2020) [22]	The Journal of Arthroplasty	Postoperative patient satisfaction	Stochastic gradient boosting	AUC: 0.81. Sensitivity: 73.0%. Specificity: 74.6%.	Demonstrated superior predictive performance than the binary logistic regression model.	Limited sample size (data from 897 cases) /  Training and testing sets were selected from the same dataset.	Identified important predictors for postoperative patient satisfaction (e.g., age).
Harris (2019) [27]	Clinical Orthopaedics and Related Research	Postoperative 30-day complications and mortality	LASSO regression	AUC: 0.72 (cardiac complications), 0.69 (mortality), 0.60 (renal complications).	Different datasets were used for initial training and testing.	Training dataset (ACS-NSQIP) does not contain complete patient medical data (e.g., comorbidities), and includes patients from a limited number of hospitals.	Developed an externally validated model using routine clinical data as predictors and could therefore potentially be used to identify high-risk patients preoperatively.