Of course, the objective is to enhance decision making, not replace clinical judgment. Ī fundamental premise of clinical risk assessment tools is that prediction based upon multiple risk factors is more accurate than prediction based upon a single risk factor, including BMD, which is consistent with the available data. Machine learning can train algorithms efficiently from large and complex data, though continuous variables such as age, BMD, or number of fall episodes generally result in a better performing algorithms than when clinical risk factors are represented as binary information. Algorithms may be designed for self-scoring by patients or they can be designed to use data stored in administrative databases or electronic patient records, where a complex algorithm may be no more difficult or time consuming to implement than a simple one. Where a tool is derived from several cohorts, such as FRAX described later, the maximum complexity also becomes limited by the extent to which a variable is registered in all cohort (e.g., femoral neck BMD) or only in some of the contributing cohorts (e.g., falling or spine BMD). The point at which this happens depends in part on the degree to which individual variables provide similar or incremental risk information. There is a law of diminishing returns in terms of the incremental value of adding further predictors to a risk algorithm once it has moved beyond a certain level of complexity. Comparison between various fracture risk assessment tools. Source: Reprinted with permission Leslie WD, Lix LM. To date, no fracture prediction tool satisfies all of these criteria. In other words, it should be accurate both in terms of the onset and offset of treatment effects. The ideal tool should also detect and report changes in risk attributable to lifestyle interventions or use of drugs with antifracture efficacy. Discrimination (the model’s ability to distinguish between individuals who do or do not experience the event of interest) and calibration (agreement between observed and predicted event rates for groups of individuals) are key aspects of predictive performance of risk algorithms. The ideal fracture risk assessment tool should be accurate, reproducible, simple, and intuitive to use, exhibit validity-discrimination with appropriate calibration in diverse populations-and lend itself to variable time scenarios and fracture types. 66.1 identifies the major steps in the development and validation of these risk estimates. Tools to estimate the risk of fracture, as opposed to screening for osteoporosis which is covered in Chapter 62, Who should be screened for osteoporosis?, have been around for almost two decades. Intuitively, it makes more sense to intervene in patients at high absolute risk than in patients whose risk of fracture is low and such an approach is also favorable in terms of health economics. Integration of clinical risk factors and measurements such as bone mineral density (BMD) into a reliable estimate of the likelihood of fracture is important for decision-making both at the societal level and for the individual patient. This results in a higher number of all fracture types, both high energy and low energy fractures, though the latter will typically lead to suspicion of osteoporosis. Osteoporosis is defined as a state of reduced bone mass and bone strength which leads to a higher than normal risk of fracture.