X-ray Based Deep Learning Artificial Intelligence Models for Predicting Proximal Junctional Kyphosis (PJK) following Adult Spinal Deformity Surgery

Introduction: Proximal junctional kyphosis (PJK) following spinal deformity surgery remains very difficult to predict. Artificial intelligence possesses the potential to extract complex features from imaging, but there is a paucity of literature exploring this topic.

Methods: 498 patients who underwent adult spinal deformity (ASD) surgery with 5 or more fusion levels from a single institution between 2015 and 2022 were included. Preoperative and initial postoperative anteroposterior (AP) and lateral x-rays were collected for all patients. A deep learning approach using convolutional neural networks (CNNs) was used to train and test models to predict PJK. Seven different models were trained and evaluated using different combinations of the four x-ray image types as data inputs for the model. The seven models were preop AP only, preop lateral only, postop AP only, postop lateral only, preoperative (preop AP and preop lateral together), initial postoperative (postop AP and postop lateral together), and combined preoperative and initial postoperative (all four image types together). Model performances were evaluated using accuracy, sensitivity, specificity, and area under the receiver operating curve (AUROC) score. Gradient class activation mapping, which highlights hotspots on each image that are important to model decision-making, was conducted to qualitatively determine which regions of each image was important for the prediction of PJK.

Results: 25 of the 498 (5.0%) patients developed PJK post-operatively. When evaluating on a test cohort, the model performances, measured by AUC, were 0.65 (preop AP), 0.54 (preop lateral), 0.71 (postop AP), 0.69 (postop lateral), 0.59 (preoperative), 0.63 (initial postoperative), and 0.82 (combined preoperative and initial postoperative). Gradient class activation mapping highlighted that in preoperative images, the model focused on the overall deformity, alignment, and balance of the patient, whereas in postoperative images, the model focused on regions where instrumentation was present, especially in the thoracic region. Models utilizing data from the initial postoperative period typically performed better. Ultimately, the model that incorporated all data together had the highest performance.

Conclusion : This is the first study to demonstrate that deep learning can conduct advanced computational analyses to detect and synthesize complex features from perioperative imaging to predict PJK.