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

Background: Poor results occasionally occur after unicompartmental knee replacement (UKR). It is often difficult, even for experienced surgeons, to determine why patients have poor outcomes from radiographs. The aim was to compare the ability of experienced surgeons and machine learning to predict whether patients had poor or excellent outcomes from radiographs.

Methods: 924 one-year anterior-posterior radiographs post-UKR were used to train a machine learning model (ResNet50v2) with a transfer learning approach based on their one-year Oxford Knee Score categories. Two experienced surgeons and the model assessed and categorised 70 radiographs (14 Poor scores; 56 Excellent scores) not used for training according to their expected outcome.

Results: The ResNet50v2 model correctly identified 71% (n = 10) of the patients with a poor score and 46 (82%) of those with an excellent score. In contrast, one surgeon could not identify patients with Poor scores (0%) and the other identified one (7%). Both misidentified 3 of those with Excellent scores. The model visualisation method suggested that estimated classifications were made from image features around the implants.

Conclusion: The results suggest that there are radiographical features that relate to poor outcomes, which the surgeons are unaware of. Those the model did not identify may have an extra-articular cause for their poor outcome. Further analysis to identify the features associated with poor outcomes could potentially suggest ways that indications or techniques could be improved so as to decrease the incidence of poor results.