Machine learning is better than surgeons at assessing unicompartmental knee replacement radiographs

Area of Science:

• Orthopedic surgery
• Artificial intelligence in medicine
• Medical imaging analysis

Background:

• Unicompartmental knee replacement (UKR) can yield suboptimal patient outcomes.
• Radiographic assessment of poor UKR results is challenging for even experienced surgeons.
• Identifying predictors of poor UKR outcomes is crucial for improving surgical success.

Purpose of the Study:

• To compare the diagnostic performance of machine learning (ML) and experienced surgeons in predicting UKR outcomes from radiographs.
• To investigate whether ML can identify subtle radiographic features associated with poor UKR results.

Main Methods:

• A ResNet50v2 ML model was trained on 924 post-UKR radiographs with known Oxford Knee Score outcomes.
• The trained ML model and two experienced orthopedic surgeons evaluated 70 unseen radiographs for predicted outcome.
• Performance was assessed by comparing predictions against actual patient outcomes.

Main Results:

• The ML model achieved 71% accuracy for poor outcomes and 82% for excellent outcomes.
• Surgeons demonstrated significantly lower accuracy, with one identifying 0% of poor outcomes.
• Model interpretability suggested classifications were based on periprosthetic features.

Conclusions:

• Machine learning models can detect radiographical indicators of poor UKR outcomes that are not apparent to human surgeons.
• Unidentified cases may have extra-articular causes for poor outcomes.
• Further research into ML-identified features could refine UKR indications and surgical techniques to minimize poor results.