A systematic review has evaluated the effectiveness of artificial intelligence (AI) in detecting and predicting vertebral fractures, adhering to rigorous methodologies aligned with Cochrane guidelines and PRISMA.

The review revealed that traditional machine learning models excel in predicting vertebral fractures, while deep learning models demonstrate superior accuracy in diagnosing various fracture types.

Timely detection and treatment of vertebral fractures pose significant challenges in healthcare, underscoring the potential role of AI and machine learning in enhancing diagnosis and prognostication.

As osteoporosis rates rise, the incidence of vertebral fractures is expected to increase, yet many fractures remain undetected, leading to higher mortality risks and chronic pain.

Globally, vertebral fractures account for approximately 8.6 million cases annually, with risk factors including inactivity, chronic conditions, smoking, and previous falls.

While AI's clinical applications have rapidly expanded in fields like dermatology and orthopaedics, its use in spinal neurosurgery remains limited.

The current diagnostic approach for vertebral fractures involves a multidisciplinary team, but it faces challenges such as inaccuracies and inefficiencies, making AI a promising alternative.

The findings suggest a need for more comprehensive datasets to enhance AI training and model performance across diverse clinical scenarios.

In total, 14,161 studies were screened for this review, resulting in 79 studies included, with 40 also featured in the meta-analysis, reflecting a strong focus on recent research in 2023.

Future research should aim to integrate the predictive strengths of traditional machine learning with the diagnostic capabilities of deep learning to improve vertebral fracture assessment.

Vertebral fractures are the most common type of fragility fractures associated with osteoporosis in the elderly, with incidence rates of 10.7 per 1000 person-years for women and 5.7 for men aged 50 and older in Europe.

The meta-analysis within the review highlighted variability in model performance, with area under the receiver operating characteristic (AUROC) scores ranging from 0.72 to 0.94, indicating that some models significantly outperform others.