Using machine learning to predict internal rotation after anatomic and reverse total shoulder arthroplasty.

Area of Science:

• Orthopedic Surgery
• Machine Learning in Medicine
• Biostatistics

Background:

• Predicting internal rotation (IR) improvement after shoulder arthroplasty, especially reverse total shoulder arthroplasty (rTSA), remains challenging due to significant patient variability.
• Anatomic total shoulder arthroplasty (aTSA) generally shows more predictable IR gains compared to rTSA.

Purpose of the Study:

• To quantitatively compare IR scores between aTSA and rTSA patients.
• To develop and validate supervised machine learning models for predicting postoperative IR at multiple time points.

Main Methods:

• Analysis of clinical data from 2270 aTSA and 4198 rTSA patients using three machine learning techniques.
• Prediction of IR scores at six postoperative intervals using full and minimal preoperative feature sets.
• Evaluation of model accuracy using Mean Absolute Error (MAE) and prediction of achieving Minimal Clinically Important Difference (MCID) and Substantial Clinical Benefit (SCB) thresholds.

Main Results:

• rTSA patients exhibited lower mean IR scores and less improvement compared to aTSA patients across all postoperative time points.
• Machine learning models using minimal preoperative data achieved high accuracy (0.92-1.18 MAE for aTSA, 1.03-1.25 MAE for rTSA) in predicting IR scores.
• Algorithms accurately identified patients likely to achieve MCID (90% for aTSA, 85% for rTSA) and SCB (85% for aTSA, 77% for rTSA) for IR improvement.

Conclusions:

• Supervised machine learning accurately predicts active internal rotation after both aTSA and rTSA using a concise set of preoperative variables.
• These predictive models can reliably identify patients who will or will not achieve clinically significant improvements in IR beyond established thresholds.