Using machine learning to predict clinical outcomes after shoulder arthroplasty with a minimal feature set

Area of Science:

• Orthopedic Surgery
• Machine Learning
• Biomedical Data Science

Background:

• Machine learning models were developed using data from 5774 shoulder arthroplasty patients.
• The study aimed to predict clinical outcomes for anatomic total shoulder arthroplasty (aTSA) and reverse total shoulder arthroplasty (rTSA).

Purpose of the Study:

• To compare the predictive accuracy of a comprehensive 291-parameter model versus a simplified 19-parameter model.
• To evaluate the efficacy of a minimal feature set as a clinical decision-support tool for shoulder arthroplasty.

Main Methods:

• XGBoost machine learning technique applied to clinical data from 2153 aTSA and 3621 rTSA patients.
• Models predicted multiple outcome measures at various postoperative time points.
• Mean Absolute Errors (MAEs) quantified prediction accuracy; patient stratification for clinical improvement was also assessed.

Main Results:

• Full and abbreviated models demonstrated comparable MAEs across all analyzed outcome measures and time points.
• Minor MAE improvements were noted when the abbreviated model included implant and native anatomy data.
• Both models effectively risk-stratified patients based on preoperative data for predicting significant clinical improvement.

Conclusions:

• Machine learning models, both full and abbreviated, achieved similar accuracy in predicting clinical outcomes post-aTSA and rTSA.
• A minimal feature set (19 preoperative inputs) shows potential as an efficient tool for improving surgical decision-making during consultations.