Table 2.

Studies evaluating the ability of machine learning algorithms to predict the outcomes of TSA

Author, year	Outcomes measured	Follow up	Most accurate algorithm	Mean absolute error or accuracy	Outcomes
Gowd et al., (2019) [27]	Postoperative complications	1 Month	N/A	N/A	Machine learning is able to predict postoperative complications in a random sample of a nationwide cohort and outperformed models by comorbidity indices alone utilizing preoperative characteristics
Kumar et al., (2020) [28]	ASES, UCLA, Constant, global shoulder function, VAS pain scores, active abduction, forward flexion, and external rotation	5 + years	Wide and Deep technique	Mean Absolute Error for Wide and Deep Technique: -ASES: ± 10.1 to 11.3 points -UCLA score: ± 2.5 to 3.4 -Constant score: ± 7.3 to 7.9 -Global shoulder function score: ± 1.0 to 1.4 -VAS pain: ± 1.2 to 1.4 -Active abduction: ± 18° to 21° -Forward elevation: ± 15° to 17° -External rotation: ± 10° to 12°	All three machine learning techniques can use preoperative data to predict clinical outcomes at multiple postoperative points after shoulder arthroplasty In addition, the models correctly identified the patients who did and did not experience clinical improvement that exceeded the MCID: 93 to 99 percent accuracy for PROMs and 85 to 94 percent accuracy for measures of pain, function, and range of motion
Kumar et al., (2021) [31]	ASES, Constant score, global shoulder function score, VAS pain scores, active abduction, forward elevation, and external rotation	5 + years	N/A	Mean Absolute Error when using 19 features: -ASES: ± 12 -Constant score: ± 9.8 -Global shoulder function score ± 1.5 -VAS pain: ± 1.4 -Active abduction: ± 21.8° -Forward elevation: ± 19.2° -External rotation: ± 12.6°	Both the full and minimal models exhibited comparable MAEs for predicting each outcome measure at each postoperative time point Additionally, both the full and abbreviated models accurately identified patients who were most at risk of having poor outcomes based on MCID thresholds, enabling risk stratification of patients using only preoperative data (full model accuracy > 82 percent vs. abbreviated model accuracy > 82 percent)
Kumar et al., (2022) [30]	Internal Rotation	5 + years	XGBoost and Wide and Deep	Mean Absolute Error when using 19 features: Wide and Deep: 3–6 months: ± 1.10 6–9 months: ± 1.16 1 year: ± 1.19 2–3 years: ± 1.07 3–5 years: ± 1.04 5 + years: ± 0.96	Active internal rotation following aTSA and rTSA may be precisely predicted at various postoperative time points using a small 19 feature set of preoperative inputs. These predictive algorithms were able to determine which patients will and won't have clinical improvement in their IR score over the MCID (90 percent accuracy for aTSA and 85 percent accuracy for rTSA)
Kumar et al., (2022) [29]	SAS score, ASES score, Constant score	5 + years	Wide and Deep technique	Mean Absolute Error when using 291 features for Wide and Deep Technique: -SAS: ± 7.56 -ASES: ± 10.68 -Constant score: ± 8.25	Although the accuracy of the three machine learning algorithms varied, they all had lower MAE than the baseline average model. Machine learning may be used to predict whether patients will see clinical improvement that is greater than the MCID (96 percent accuracy for both a TSA and rTSA)
Lopez et al., (2021) [33]	Non-home discharge and 30-day postoperative complication rates	1 Month	Both had similar accuracy, but the artificial Neural Network had better discriminative ability	Accuracy for Artificial Neural Network 30-day Postoperative Complication rate Accuracy: -Boosted Decision Tree: 95.5% -Artificial Neural Network: 92.5%	Both Boosted decision tree model and Artificial Neural Networks has a greater than 90% accuracy in predicting 30-day postoperative complications
Lopez et al., (2022) [34]	Prolonged operative time and 30-day postoperative complication rates	1 Month	Artificial Neural Network	Accuracy for Artificial Neural Network 30-day Postoperative Complication rate Accuracy: -Boosted Decision Tree: 95.5% -Artificial Neural Network: 92.5%	Both Boosted decision tree model and Artificial Neural Networks has a greater than 90% accuracy in predicting 30-day postoperative complications
McLendon et al., (2021) [35]	ASES	2 years	N/A	Accuracy of predicting different improvement levels for model 1: - ≤ 28 points: 94% -29 to 55 points: 95% - > 55 points: 94%	Machine learning can reliably predict the extent of improvement following glenohumeral OA shoulder arthroplasty
Polce et al., (2020) [36]	Patient satisfaction	2 years	Support vector machine	N/A	The Support vector machine model demonstrated excellent discrimination and adequate calibration for predicting satisfaction following TSA

TSA Total Shoulder Arthroplasty, ASES American Shoulder and Elbow Surgeons score, UCLA University of California, Los Angeles Score, VAS Visual Analog Scale, MCID Minimal Clinically Important Differences