Machine Learning Solutions for Osteoporosis – A Review

Authors: Julien Smets et al. (2021)

Background Information:

Osteoporosis is a common bone disease characterized by weakened bones and a high risk of fractures, affecting millions worldwide. Traditionally, doctors assess bone health using bone mineral density (BMD) scans like DXA, sometimes combined with clinical tools like FRAX to estimate fracture risk. However, these methods can't fully capture complex patterns in imaging and patient data. Recently, Artificial Intelligence, particularly Machine Learning (ML), has offered the potential to identify hidden patterns in large and dense datasets, which could improve osteoporosis screening, fracture detection, and risk prediction.

Purpose of the Study:

This review set out to examine how ML techniques have been applied in osteoporosis care. The authors reviewed 89 studies and grouped them into four main categories: measuring bone properties, diagnosing osteoporosis, detecting fractures in images, and predicting future fracture risk. Their goal was to evaluate the current state of ML applications, identify common pitfalls, and highlight the most promising areas for clinical use.

Methods and Data Analysis:

The researchers conducted a systematic search in PubMed and Web of Science to identify studies published between 2015 and 2020. They evaluated the methodological quality of each study using a 12-point checklist (based on MI-CLAIM standards) that considered factors like model choice, data splitting, validation, and reporting transparency. Instead of combining results quantitatively, they qualitatively summarized findings across the four application areas: bone assessment, osteoporosis classification, fracture detection, and fracture risk prediction.

Key Findings and Conclusions:

The review found that ML applications in osteoporosis are promising but often inconsistent. Many studies showed good accuracy in tasks like osteoporosis detection (mean AUC ~0.90) and fracture classification, especially using image-based ML models. However, key issues were incomplete reporting, risk of overfitting, inadequate test validation, and the lack of external datasets. The most encouraging progress was seen in image-based opportunistic diagnoses—using routine scans to detect bone weakness or fractures without extra testing. Overall, the authors concluded that ML holds real potential for improving osteoporosis care, but current studies need better design and stricter validation to ensure reliable and generalizable results.

Applications & Limitations:

ML models could one day enhance osteoporosis care by enabling automated detection of low bone density or fractures on common imaging (e.g., CT, X-ray), and by accurately predicting who’s at high risk—possibly better than existing tools like FRAX. Such innovations may allow earlier intervention and reduce missed diagnoses. However, most research remains at early stages, hampered by methodological issues: small or homogeneous datasets, lack of external testing, poor reporting, and risk of overfitting. To progress, future studies should follow robust standards, validate in diverse populations, and consistently share data and code so these tools can be trusted in clinical settings.

Return