Abstract
BACKGROUND: Ossification of the posterior longitudinal ligament (OPLL) is a significant contributor for unintentional durotomy following anterior spinal surgery, neural compression, and cervical myelopathy. While traditional diagnostic methods like plain radiography are commonly used, they may yield false negatives. The diagnostic accuracy and reliability of artificial intelligence methods for detecting this condition remain largely unexplored. This study aimed to systematically evaluate the performance of deep learning models (DLMs) in diagnosing and predicting cervical OPLL. METHODS: This systematic review assesses the utilization of DLMs in diagnosing and predicting OPLL. Inclusion criteria were defined as the use of DLM for the diagnosis and prediction of cervical OPLL in adult patients. Databases included PubMed, Google Scholar, Cochrane Library, ScienceDirect, and BASE. The risk of bias was assessed using the QUADAS-2 tool. RESULTS: Seven studies with a pooled sample size of 3,373 patients were included. The pooled accuracy, area under the curve, sensitivity, and accuracy are 0.93, 0.92, 0.88, and 0.9. DLM demonstrated superior diagnostic performance, outperforming human comparator groups in terms of sensitivity (0.86 vs. 0.77), specificity (0.98 vs. 0.74), and accuracy (0.89 vs. 0.76). The meta-analysis with a pooled sample size of 1,016 patients revealed the highest proportion of right-identified OPLL subtypes in the mixed- and continuous subtypes (0.93 and 0.87). Accuracy and sensitivity of DLM were higher in the upper compared to the lower cervical spine. CONCLUSIONS: Despite limitations in methodological variations and deep learning challenges, the findings support integrating these models into diagnostic protocols. Their robust performance suggests potential value in clinical practice, offering improved diagnostic accuracy and enhanced subtype differentiation.