Deep learning-based multimodal image analysis predicts bone cement leakage during percutaneous kyphoplasty: protocol for model development, and validation by prospective and external datasets.

Background: Bone cement leakage (BCL) is one of the most prevalent complications of percutaneous kyphoplasty (PKP) for treating osteoporotic vertebral compression fracture (OVCF), which may result in severe secondary complications and poor outcomes. Previous studies employed several traditional machine learning (ML) models to predict BCL preoperatively, but effective and intelligent methods to bridge the distance between current models and real-life clinical applications remain lacking. Methods: We will develop a deep learning (DL)-based prediction model that directly analyzes preoperative computed tomography (CT) and magnetic resonance imaging (MRI) of patients with OVCF to accurately predict BCL occurrence and classification during PKP. This retrospective study includes a retrospective internal dataset for DL model training and validation, a prospective internal dataset, and a cross-center external dataset for model testing. We will evaluate not only model's predictive performance, but also its reliability by calculating its consistency with reference standards and comparing it with that of clinician prediction. Discussion: The model holds an imperative clinical significance. Clinicians can formulate more targeted treatment strategies to minimize the incidence of BCL, thereby improving clinical outcomes by preoperatively identifying patients at high risk for each BCL subtype. In particular, the model holds great potential to be extended and applied in remote areas where medical resources are relatively scarce so that more patients can benefit from quality perioperative evaluation and management strategies. Moreover, the model will efficiently promote information sharing and decision-making between clinicians and patients, thereby increasing the overall quality of healthcare services.