What a FRAX Fracture Risk Score Can and Cannot Tell You

A FRAX score estimates your 10-year probability of two things: a hip fracture, and a major osteoporotic fracture (hip, spine, forearm, or shoulder). It builds that estimate from your age, sex, weight, height, and a short list of clinical risk factors, and it can run with or without a bone-density measurement plugged in. What it does well is sort a population by average risk. What it does less well, and what the marketing around any risk calculator tends to obscure, is tell one specific person what will happen to them. Adding a bone-density scan improves the tool's accuracy only modestly.

This article is educational and is not medical advice.

What FRAX actually computes

FRAX is a statistical model. It was developed from large international cohorts and it outputs, per the tool's own documentation, the 10-year probability of hip fracture and of major osteoporotic fracture for people aged roughly 40 to 90. The inputs are deliberately simple: age, sex, a prior fragility fracture, parental hip fracture, current smoking, glucocorticoid use, rheumatoid arthritis, secondary causes of osteoporosis, alcohol intake, and body size. You can optionally add a femoral-neck bone mineral density (BMD) value from a DXA scan.

That design is a feature. Because most of the inputs are answerable in a brief conversation, FRAX can flag who might benefit from a scan before anyone is scanned. The 2024 draft and 2025 final recommendations from the US Preventive Services Task Force lean on exactly this logic: the Task Force gives a B recommendation to bone-density screening for women 65 and older, and a B recommendation for postmenopausal women under 65 whose risk, estimated by clinical risk assessment, is already elevated. For men, the Task Force issued an I statement, meaning the evidence is insufficient to weigh benefits against harms. The final statement describes the risk assessment in general terms, while the draft names tools such as FRAX as ways to do that upstream sorting.

Why bone density adds less than you would expect

Here is the counterintuitive part. Bone density is the thing most people associate with fracture risk, yet feeding a BMD value into FRAX sharpens the prediction only a little.

The evidence for this is consistent across cohorts. In the Study of Osteoporotic Fractures analysis of the WHO FRAX models, the model with BMD predicted major osteoporotic fractures with an area under the curve (AUC) of roughly 0.61 to 0.64 depending on the woman's baseline bone status, and predicted hip fractures somewhat better, with AUCs around 0.62 to 0.78. Crucially, the authors reported that prediction with the full FRAX model was similar to much simpler models. Age plus a history of prior fracture captured most of the signal; the additional clinical risk factors, and even BMD itself, added modest incremental discrimination. A 2019 systematic review and meta-analysis of fracture-prediction tools reached a compatible conclusion, reporting that for osteoporotic fracture FRAX discriminated with a pooled AUC of about 0.72 when BMD was included and about 0.69 when it was not. That gap is real but small, which is the whole point: the scan nudges the ranking rather than transforming it.

Two lessons fall out of this. First, bone density is a strong risk factor but not a deterministic one; many fractures happen in people whose density is not in the osteoporotic range. Second, the marginal value of a test depends on what you already know. If age and fracture history have already placed someone in a high-risk band, the scan may confirm rather than overturn the picture.

How to read an AUC without fooling yourself

AUC, also called the C-statistic, is the single number most often cited to defend a risk calculator, so it is useful to know how to read it. AUC is the probability that the model assigns a higher risk score to a person who will fracture than to a person who will not, when you draw one of each at random. An AUC of 0.5 is a coin flip. An AUC of 1.0 is perfect ranking.

Most FRAX estimates land in the 0.6 to 0.8 range. That is genuinely useful for population-level decisions and unimpressive for individual prophecy. An AUC of 0.70 means the model ranks the eventual fracture case above the non-case 70 percent of the time, which also means it gets the pairwise order wrong nearly a third of the time. A tool can have a respectable AUC and still misclassify many individuals, because ranking a population correctly on average is a much easier task than calling one person's fate.

Three appraiser's habits follow from this:

Ask what the AUC is being compared against An AUC of 0.75 sounds strong until you learn a two-variable model reaches 0.73. The right question is not "is this number high" but "how much does it beat the cheapest available alternative." In fracture prediction, the honest answer is often "a little."

Separate discrimination from calibration AUC measures ranking only. It says nothing about whether the predicted 10-year risk of 12 percent corresponds to a real-world 12 percent. A model can discriminate acceptably yet systematically over- or under-predict, which is why the FRAX literature also flags over-prediction in some groups. A calculator that is well calibrated in the population it was built on can drift when applied elsewhere.

Treat the output as a probability, not a verdict A 10-year major-fracture probability is a base rate for people like you, not a countdown clock. It should open a conversation about whether further testing or treatment changes your expected outcome, weighed against the harms of acting.

What the score is good for

None of this makes FRAX a poor tool. It makes it a tool with a defined job. FRAX is well suited to triage, to standardizing who gets referred for a scan, and to framing a shared decision about whether the benefit of intervention outweighs its burdens. It is poorly suited to being read as destiny, to comparing two individuals whose scores differ by a few points, or to replacing the clinical judgment that the USPSTF itself says must accompany any risk estimate. The number is an input to a decision, and the quality of the decision depends on how carefully the number is interpreted.