How Clinical Risk Changes the Calculation
In osteoporosis assessment, the difficult work begins after the bone-density measurement is available. This chapter focuses on the clinical variables and risk tools that surround a T-score: what each tool is designed to predict, why adding BMD can improve an estimate, and why a more elaborate number is still conditional on its inputs. That distinction helps readers use risk estimates as evidence for a decision rather than mistaking them for a diagnosis or an automatic treatment order.
A risk tool must answer a specific question
A risk tool is not simply a second kind of T-score. Its purpose depends on the outcome it was built to identify or predict. The USPSTF separates instruments designed to identify likely osteoporosis from instruments designed to estimate future fracture risk. That distinction is easy to overlook because both kinds of tool may be described with the general word “risk.” They do not, however, answer the same question.
Tools such as the Osteoporosis Self-Assessment Tool and the Osteoporosis Risk Assessment Instrument generally use a relatively small number of inputs to identify people who are more likely to have osteoporosis. Their practical role can be to help decide who should receive a bone-density test. They are not designed to replace DXA or to provide the same kind of 10-year fracture probability that a fracture-prediction model provides.
FRAX, the Fracture Risk Calculator and the Garvan Fracture Risk Calculator have a different job. They are designed to estimate future fracture risk using demographic and clinical information, with some able to incorporate BMD. FRAX estimates the 10-year probability of a hip fracture or a major osteoporotic fracture for people aged 40 to 90. Its major osteoporotic-fracture category includes fractures of the hip, spine, wrist or shoulder. Depending on the version and the assessment, FRAX can use clinical factors alone or combine them with BMD measured at the femoral neck.
The time horizon is part of the output. “Ten-year probability” is not the same as “has osteoporosis,” “will fracture eventually,” or “needs medication now.” A prediction is meaningful only when the outcome, time period and population are clear. A tool that identifies likely low BMD can be useful before DXA; a tool that estimates fracture probability can help put a measured BMD result into a broader context. Neither tool, simply by producing a score, settles every decision that follows.
The inputs also reveal what the model is trying to capture. Clinical assessment may include age, weight or body size, a previous fracture, a parental hip fracture, smoking, alcohol use, glucocorticoid exposure, rheumatoid arthritis or a condition classified as secondary osteoporosis. The Canadian quick-reference guide similarly asks clinicians to look for prior fragility fracture, repeated falls, low body mass index, parental hip fracture, smoking, alcohol intake, secondary osteoporosis and prolonged glucocorticoid use. These are not decorative additions to a density report. They connect the measurement to the circumstances in which a fracture is more likely to occur.
But adding a factor does not make the model a complete representation of the person. Each variable is simplified, defined in a particular way and included because the developers believed it improved prediction in the data used to build or validate the tool. A clinical history can contain information that the form does not ask for. A risk estimate therefore describes a person through the model’s chosen lens, not through every fact that might matter.
That is the first discipline of using a risk tool: ask what it was built to do before asking what its number means. A score intended to identify osteoporosis should not be treated as a fracture probability. A fracture probability should not be treated as proof of osteoporosis. And either result should not be treated as a treatment decision without considering the evidence for acting on it.
This relationship map is intentionally not a straight line from a person to a prescription. Clinical factors and BMD are information about the person. FRAX is one tool that can use those inputs, including femoral-neck BMD, to produce an estimate. The estimate informs a clinical decision; it does not own that decision. The distinction between an input, an estimate and an action is where much of the safety of the pathway resides.
Adding BMD improves a model—but does not make it neutral
The reason to combine clinical risk information with BMD is not that either source of information is worthless alone. It is that they carry different signals. BMD describes density at a measured skeletal site. Clinical factors describe characteristics and exposures associated with fractures. A tool can use both to rank risk more effectively than either signal alone in many of the studies reviewed by the USPSTF.
The evidence is not merely theoretical. Across studies of fracture-prediction instruments, findings were heterogeneous, but the predictive accuracy of tools that could accept BMD—FRAX, the Fracture Risk Calculator and Garvan—often improved when BMD was included. For FRAX in women, the reviewed systematic reviews reported AUCs of 0.65 to 0.67 for 10-year major osteoporotic-fracture risk without BMD and 0.67 to 0.71 with BMD. For hip-fracture risk, the ranges were 0.74 to 0.77 without BMD and 0.76 to 0.79 with BMD.
These figures support a modest conclusion. BMD can add information to a clinical risk model. They do not support the stronger conclusion that the combined output is a direct readout of an individual’s future. The AUC still describes discrimination across studied groups, not a guarantee for the person entering the calculator. The studies also vary in their populations, outcomes, follow-up and thresholds, so the ranges should not be treated as universal performance specifications.
A combined estimate is conditional in several ways. It is conditional on the selected inputs being accurate. It is conditional on the model’s definitions matching the person’s circumstances. It is conditional on the development and validation populations being relevant to the person being assessed. And it is conditional on the decision being made: a tool can predict an outcome without proving that acting on the tool’s output improves that outcome.
The last point is especially important. The USPSTF reports that no published study it identified was designed to evaluate a treatment strategy based on fracture risk from FRAX alone. This does not mean FRAX is useless. It means that predictive accuracy and treatment-strategy effectiveness are different kinds of evidence. A model may rank people by risk, while a separate clinical trial is needed to show that using that ranking to select treatment produces more benefit than harm.
A percentage can hide this distinction because it looks final. Suppose a calculator reports a 10-year probability. The number may be helpful for comparing a person’s estimated risk with the context of a guideline or a shared decision. But it is not itself a trial result showing what will happen if treatment is started, and it is not a diagnosis produced by the calculator. The decision still belongs to the clinical pathway that defines the population, the available interventions, expected benefits, potential harms and the person’s situation.
The assumptions become more consequential when a model leaves out factors that matter. The USPSTF identifies several limitations of FRAX. In the United States, race-specific versions calibrated to observed fracture-incidence data produce lower predicted risk for Asian, Black and Hispanic people than for White people with the same age, sex, weight, BMD and listed clinical risk factors. The statement also notes uncertainty about which version to use for people who are multiracial or who immigrated from another country.
This is not a claim that a person’s biology can be read from a racial label. It is a warning about how a calibration choice can change who crosses a threshold when all of the listed clinical inputs appear identical. If a model is trained on different background rates and then paired with a fixed action threshold, the same apparent profile can lead to different estimated risks. The result may look objective because it is numerical, while still depending on a contestable population assumption.
Other limitations are less visible in the final output. FRAX treats glucocorticoid exposure and alcohol use as binary inputs rather than fully quantified doses. It does not include lumbar-spine BMD or trabecular bone score, and it does not collect information about falls or frailty. The USPSTF also points to the use of older cohort studies for race-specific fracture incidence, mortality estimates not updated since 2004, and the omission of medical conditions such as diabetes that may signal increased risk. These omissions do not prove that every estimate is wrong. They show why an estimate can understate risk when important information is outside the model.
This matters most when a person’s circumstances are unlike the average case represented in the data. Consider two hypothetical people with the same T-score and the same listed FRAX inputs. One may have a history of frequent falls that the calculator does not record; the other may not. The resulting estimate could be identical even though the clinical context is not. The appropriate lesson is not to invent a corrected percentage. It is to recognize that the shared output is conditional on the information the tool can see.
A similar issue arises when a medical condition or medication changes the pathway. A risk model may include a broad category such as secondary osteoporosis or glucocorticoid use, but a single checkbox cannot express every underlying disease, dose, duration or treatment history. The USPSTF’s screening recommendation excludes people with secondary osteoporosis due to specified conditions or chronic use of medications associated with bone loss. That boundary is a reminder that a screening calculator designed for one population should not be assumed to cover every cause of low bone density.
The Canadian pathway makes the relationship between tool and measurement explicit in a different way. Osteoporosis Canada’s quick-reference guide directs clinicians to calculate 10-year fracture risk with BMD using FRAX as preferred or CAROC when considering pharmacotherapy. The guide also begins with history and physical assessment, including risk factors for low BMD, fractures and falls. In this structure, the BMD result is not removed from the decision; it is brought into a risk calculation after the clinical context has been considered.
That does not make the Canadian approach a universal correction for every limitation in FRAX or establish that all settings should use the same sequence. It demonstrates a general principle: the meaning of a number depends on where it sits in the decision system. A tool used to select people for DXA serves a different purpose from a tool used to estimate treatment-related fracture risk after BMD is known. A result used in a screening program serves a different purpose from a result interpreted for someone with a prior fragility fracture or a secondary cause of bone loss.
The practical reading rule is therefore simple but demanding. First identify the outcome: osteoporosis classification, hip fracture, major osteoporotic fracture or something else. Then identify the time horizon, inputs and population behind the estimate. Next ask which important risk factors are missing or simplified. Only after those questions should the number be connected to evidence about screening or treatment.
A T-score remains part of that calculation because BMD adds information. Clinical risk changes the calculation because fracture is not a property of density alone. The combined result becomes more useful when its purpose and assumptions are visible, not when its decimal places are mistaken for certainty. The next question is how health systems turn this layered assessment into screening and management pathways, and what the clinical evidence says those pathways actually achieve.