A new study has revealed that machine learning models can significantly outperform experienced surgeons in predicting patient outcomes after partial knee replacement.
Unicompartmental (partial) knee replacement (UKR) is a common procedure designed to relieve pain in patients with knee osteoarthritis. However, some patients experience poor outcomes, and it can be challenging for surgeons to determine why these issues arise.
A new study by scientists at the University of Oxford’s Nuffield Department of Orthopaedics, Rheumatology and Musculoskeletal Sciences (NDORMS), supported by the NIHR Oxford Biomedical Research Centre, found that a machine learning model can better predict which patients will have poor outcomes after undergoing UKR surgery, compared to experienced orthopaedic surgeons. The research was published in the journal The Knee.
“It is often difficult, even for experienced surgeons, to determine why some patients have poor outcomes after partial knee replacements just by looking at the X-rays,” said lead author Dr Jack Tu, Research Fellow in Clinical Biomechanics at NDORMS.
“So we designed a study to understand whether AI had the potential to enhance surgical decision-making and improve patient care.”
The team trained a machine learning model to identify patterns in images that are associated with poor outcomes, and most of these images look normal to surgeons. They then analysed more than 900 X-rays taken one year after surgery and compared the ability of seasoned surgeons to predict outcomes with that of the machine learning model.
The results were significant. The machine learning model accurately identified 71 percent of patients with poor outcomes, while the surgeons struggled, identifying between zero and seven percent of possible issues.
This new approach using AI is different from how machine learning is typically applied in medicine. Rather than just replicating tasks that doctors can already do, the model was able to identify previously unknown visual markers that could indicate complications.
“If this technology was rolled out more widely using routine joint replacement registry data, it has the potential to significantly improve the results of partial knee replacement surgery,” Dr Tu added.
“By pinpointing the specific imaging features that lead to bad outcomes, we may be able to learn from the result which may eventually lead us to modify surgical techniques or implant designs to reduce the risk of patient dissatisfaction.”
The researchers plan to further analyse the X-ray features the AI model used to make its predictions, which could provide surgeons with new insights to optimise patient selection and surgical technique for partial knee replacements.