← Back to Home

Chest X-rays Can Also Signal Fracture Risk: AI Outperforms FRAX in External Validation Involving More Than 80,000 People

Chest images originally taken for other clinical purposes may offer clues to bone fragility. A new study shows that AI outperforms FRAX incorporating bone mineral density in distinguishing fracture risk, but another validation hurdle remains before it can be used in the general patient population.

By SURL BioNews

One challenge in fracture prevention is identifying people who truly need further assessment before a bone breaks. A study published in NEJM AI turned to existing chest X-rays: could a deep learning model read these images for the risk of future major osteoporotic fractures? In external validation involving 85,284 patients at two independent institutions, the model outperformed FRAX, the fracture risk assessment tool incorporating bone mineral density data, in distinguishing higher from lower risk.

According to the paper’s abstract indexed by Life Science Network and a summary by AgingResearch.News, the research team first developed the model using data from 53,109 adults at one medical center, then validated it at two other institutions. All participants had undergone bone mineral density testing and had a chest X-ray within one year before or after the test. The validation population had a mean age of 58.6 years, 68.9% were women, and follow-up lasted 77 months. The study aimed to predict subsequent fractures, rather than simply identify low bone mineral density on images.

The model’s AUROC for 10-year fracture prediction was 0.855, higher than the 0.785 achieved by FRAX with bone mineral density. Its C-index, another measure used to assess the ranking of event risk, was also higher, at 0.843 versus 0.779. Adding FRAX to the imaging model did not substantially improve risk discrimination, indicating that the additional benefit of FRAX was limited in this dataset.

These figures require careful interpretation. AUROC and C-index primarily reflect whether a model can rank people more likely to sustain a fracture ahead of others; 0.855 cannot be read directly as “85.5% prediction accuracy.” Whether the model can accurately estimate an individual’s actual probability of fracture, and what threshold would be appropriate for arranging further testing, are separate questions.

The study also focused on people with osteopenia—those whose bone mineral density is below normal but has not reached the osteoporosis range. The paper’s abstract notes that a substantial share of the fracture burden occurs in this range, and the imaging model also outperformed FRAX in risk discrimination within this group. This suggests that existing chest images may supplement information not fully captured by bone mineral density assessment. However, the available abstract does not provide enough detail about which imaging features the model uses to make its judgments.

The study’s most important limitation lies in its enrollment criteria: everyone had already been referred for bone mineral density testing, so the population cannot directly represent patients undergoing chest X-rays in general. External validation at two institutions strengthens the credibility of the findings, but it still does not answer how many previously missed high-risk individuals the model would identify in routine clinical workflows, or how many people it would lead to undergo additional assessment.

The potential of this work is to extract an additional layer of fracture-risk clues from images already taken, without requiring another imaging examination. Before practical implementation, the next step remains prospective evaluation in populations not selected through referral for bone mineral density testing, to confirm performance in estimating absolute risk and identifying high-risk individuals. The current findings support further validation of this approach, but are not yet sufficient to replace existing assessments or determine treatment.

References

  1. NEJM AI abstract via Life Science Network
  2. AgingResearch.News