Summary
Artificial Intelligence is able to accurately predict patient outcomes 2-years post-operatively using intraoperative arthroscopic images of the cotyloid fossa for patients undergoing hip arthroscopy.
Abstract
Introduction
Recent advancements in artificial intelligence (AI) and image recognition have greatly impacted various medical fields. Despite these advancements, challenges related to data quality, model validation, and clinical integration persist. Previous studies have demonstrated AI’s capability in predicting outcomes, however, no studies have yet evaluated this technology on intra-operative hip arthroscopy images. This study aims to evaluate the efficacy of AI in predicting patient reported outcomes (PROs) and cotyloid synovitis in patients undergoing hip arthroscopy using intraoperative arthroscopic images to train and test AI models.
Methods
This retrospective cohort study included patients aged 18-80 who underwent hip arthroscopy for femoroacetabular impingement (FAI) between January 1, 2010, and May 1, 2022, by the Principal Investigator (MJP). Patients with incomplete follow-up data or insufficient intraoperative images were excluded. The primary outcomes were preoperative and postoperative Harris Hip Score (HHS), HHS PASS, Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) scores, WOMAC MCID, and the presence of cotyloid synovitis determined at the time of surgery by the senior surgeon. To facilitate accurate model training, the study focused specifically on intraoperative arthroscopic images of the cotyloid fossa. A preliminary model, EfficientNetV2-L, was pretrained on a small set of manually screened images, achieving a 94% accuracy in identifying cotyloid fossa images. Using this model, over 26,000 intraoperative images were screened to identify 1,750 high-quality cotyloid fossa images, which were confirmed through manual review. The main model, EfficientNetV2-S, was then trained on a portion of these images (Cohort A) and tested on the remaining images (Cohort B) for different outcomes. The models’ predictions were compared against actual postoperative outcomes to assess accuracy.
Results
The study included 1,750 images from 742 cases. Overall, EfficientNetV2-S demonstrated strong performance in predicting HHS PASS and HHS MCID, and PRO quartiles. For predicting HHS PASS at two years post-operatively, the model demonstrated an accuracy of 0.84 and AUC of 0.88, with a sensitivity of 0.97 and specificity of 0.70. Additionally, the model predicted achievement of HHS MCID at two years post-operatively with an accuracy of 0.82 and AUC of 0.91, and sensitivity and specificity of 0.94 and 0.74 respectively. The model showed consistent accuracy in predicting the presence of cotyloid synovitis with an accuracy of 0.89, AUC of 0.97, and sensitivity of 0.89 and specificity of 0.94. Furthermore, the model was accurately able to predict HHS quartiles, with an accuracy of 0.90, 0.84, 0.83, and 0.63 for the first, second, third and fourth quartiles respectively at 2 years post-operatively. The model demonstrated an accuracy of 0.85, 0.86, 0.78, and 0.57 for predicting the first, second, third, and fourth quartiles for WOMAC 2-years postoperatively.
Conclusion
This study demonstrates the potential of AI and machine learning in using intraoperative arthroscopic images to predict post-surgical outcomes and the presence of cotyloid synovitis in hip arthroscopy. Additional applications of this technology include integration of AI into current operating room technology as this would allow surgeons to use intraoperative arthroscopic photos at the time of surgery to provide real-time feedback and guide intraoperative clinical decision making.