OTSR - 2026-10-17 - Journal Article
Predicting Postoperative Outcomes in Small Rotator Cuff Tears: An AI-Based Approach to Long Head of the Biceps Management.
Benhenneda R, Brouard T, Samargandi R, Charousset C, Berhouet J
Topics
Key Takeaway
A hybrid CNN-MLP deep learning model predicted 24-month Constant scores with 91.4% accuracy overall, but accuracy dropped to 59.2% in patients with preserved healthy LHB who developed poor outcomes (Constant score <70).
Summary Depth
Choose how much analysis to show on this article page.
Summary
This prospective multicenter study asked whether an AI model combining arthroscopic image analysis (CNN) and clinical/imaging data (MLP) could predict 24-month Constant scores to guide LHB preservation versus tenotomy in small rotator cuff tears with macroscopically healthy LHB. The model achieved 91.4% overall accuracy with a mean prediction error of 6.04 points, and 93% accuracy in the preserved-healthy-LHB subgroup. However, in the failed-preservation subgroup, accuracy fell to 59.2% with a mean deviation of 18 points, and the model incorrectly recommended preservation in 8 cases.
Key Limitation
The absence of a concurrent tenotomy/tenodesis control group means the model predicts outcomes of preservation but cannot inform whether intervention would have produced superior results in the 8 misclassified cases.
Original Abstract
INTRODUCTION
The management of the long head of the biceps (LHB) tendon in small rotator cuff tears remains controversial, particularly when the tendon appears macroscopically healthy. While prophylactic tenotomy or tenodesis is often recommended for pathological LHB, the decision-making process for an apparently normal LHB remains unclear. This study aimed to develop and evaluate an AI-based model for predicting postoperative outcomes in patients with small rotator cuff tears and a seemingly healthy LHB, assisting in the decision between preservation and tenotomy.
METHODS
This was a prospective, multicenter, single-blind study including 199 patients with small rotator cuff tears. Preoperative clinical, imaging, and arthroscopic data were collected. A deep learning model, integrating a convolutional neural network (CNN) for arthroscopic image analysis and a multilayer perceptron (MLP) for clinical and imaging data, was developed to predict the Constant score at 24 months follow-up. The model's predictions were compared with traditional statistical analyses, and subgroup analyses were performed on patients with a preserved healthy LHB and those with a poor functional outcome, defined by a Constant score <70 at follow-up.
RESULTS
The AI model accurately predicted the Constant score at 24 months in 91.4% of cases, with a mean prediction error of 6.04 points. In the "preserved healthy LHB" subgroup, the model achieved 93% accuracy, with a mean deviation of 5.45 points. However, in the "failed healthy LHB preserved" subgroup, accuracy dropped to 59.2%, with a mean deviation of 18 points. The AI model overestimated outcomes in 8 cases (leading to incorrect preservation of the LHB) and underestimated in 6 cases.
CONCLUSION
An AI model combining arthroscopic imaging and clinical data can predict postoperative Constant scores with high accuracy. However, its performance is limited in predicting failures, particularly in patients with a preserved LHB who later develop poor outcomes. While AI may complement conventional statistical methods in clinical decision-making, further validation with larger datasets and refined variables is necessary to enhance its reliability and clinical relevance.
LEVEL OF EVIDENCE
III; Diagnostic Study.