Summary

ML models accurately predicted functional outcomes 2-3 years after shoulder arthroplasty. Implementing ML models in clinical evaluation and preoperative decision-making can help stratify the risk of patients with poor outcomes after shoulder arthroplasty

Abstract

Background

The use of machine learning (ML) algorithms in disease classification and outcomes prediction is growing. ML techniques can represent valuable tools for informed decision-making through several prediction models. In the setting of shoulder arthroplasty, recent reports showed that ML predictive models can accurately predict postoperative and functional outcomes, which can help stratify patients according to their preoperative characteristics. In this systematic review, we aimed to assess the utility of ML models in predicting the clinical outcomes after shoulder arthroplasty.

Methods

We searched PubMed, Scopus, Embase, and Cochrane databases of the diagnostic accuracy studies assessing the predictive value of ML models in patients undergoing shoulder arthroplasty from January 2010 to May 2022. The diagnostic accuracy measures were extracted in the form of the area under the curve (AUC).

Results

The present systematic review retrieved four studies assessing patients who underwent either anatomic total shoulder arthroplasty (n = 4895 patients) or reverse total shoulder arthroplasty (n = 10618 patients). All included studies used extreme gradient boosting (XGBoost) and linear regression to develop the ML models. Besides, the Wide and Deep technique was used in one study. The included studies utilised a full range of baseline variables to build the predictive models. In addition, two studies developed abstracted models by omitting preoperative functional scores and morphological features. The following outcomes were assessed: American Shoulder and Elbow Surgeons, pain scores, internal rotation score, and postoperative complications. The full XGBoost models showed high accuracy in predicting ASES (77-94%), Internal rotation score (85-90%), postoperative complications (68.1%), and patient-reported outcome measures.

Conclusions

In conclusion, ML models accurately predicted functional outcomes 2-3 years after shoulder arthroplasty. Both full and abstracted models achieved high accuracy in the prediction of global functional scores, pain scores, and rotation. Nonetheless, the current literature also suggests full ML models have higher accuracy than abstracted models in predicting clinical outcomes. Such findings highlight that implementing ML models in clinical evaluation and preoperative decision-making can help stratify the risk of patients with poor outcomes after shoulder arthroplasty.