The natural history of osteoarthritis is slow, progressive, and often unpredictable. That sluggishness stretches the window of care over many years, raising important but often neglected questions: How does someone’s physical activity level change over time? Does this put them at higher risk for disease progression? Are past physical activity habits more important than present ones? Current guidelines provide sweeping recommendations promoting physical activity, citing studies showing pain relief and improved function1. However, many of these studies are cross-sectional and fail to capture the nuance of time-dependent change. Cai et al. take a creative and commendable approach to this problem. They use joint models, a statistical technique, to analyze repeated-measurement data (i.e., physical activity) from the Osteoarthritis Initiative (OAI) and time-to-event outcomes (i.e., joint space loss) to illustrate the subtle changes in physical activity associated with joint space loss according to Kellgren-Lawrence (KL) grade. In summary, Cai et al. identify moderate physical activity as protective in KL grade 2 and higher current physical activity levels as a risk factor in KL grade 3. Structural progression may partly be stage-dependent. Clearly, the authors are not suggesting a complete cessation of activity at KL grade 3; rather, the dynamic methodology shows that activity recommendations need to be personalized to their goals or general health. How do we get to a place of “personalized care” that is simultaneously data-driven? The answers are not likely to be found in legacy data sets such as the OAI. When these data sets are curated, they are agnostic to any specific hypothesis, stitching together a patchwork of relevant and less relevant information with the aim of crowdsourcing novel insights. Consequently, they pass forward data from status-quo questionnaires like the Physical Activity Scale for the Elderly (PASE) as a loose proxy for longitudinal behaviors that researchers then funnel into their outcome-prediction analysis2. Although Cai et al. attempt to stratify “moderate” from “high” by offering a range of step counts (with moderate being walking approximately 6,000 to 9,000 steps per day and high being walking approximately ≥10,000 steps per day), the PASE alone remains a poor correlator for such metrics and cannot account adequately for the intensity of activity3. Moreover, if we ask patients what they think “moderate” activity is, more likely than not, we will receive variable answers depending on to whom or even when the question is posed4. This subjectivity is the frail foundation upon which survey questionnaires like the PASE rest and what every subsequent analysis propagates. The continued reliance on measures such as the PASE in osteoarthritis epidemiology reflects convenience rather than precision, blurring distinctions between biomechanically different activities and entirely missing the fluctuating nature of people’s habits. Approaches such as joint models point toward something crucial. We need to capture dynamics rather than snapshots; that much is true. However, the robust work by Cai et al. is not meant to inspire a rehashing of new statistical tools used on the same old data set and hoping for actionable conclusions to emerge from swapping a covariate or an outcome variable. What both clinicians and patients want is sharper, actionable guidance: how much to walk, whether cycling is safer than jogging, and what “moderate” really means in practice for them. To get there, we need more than clever modeling. We need data that match the ambitious questions that we are asking. Reassuringly, researchers in orthopaedics have already begun to take these strides forward. The APPROACH project represents such an advancement. Researchers for this project spent considerable effort on the recruitment of clinical trial participants using multimodal data and advanced algorithms for the selection of cohorts to analyze prospectively5. By painstakingly designing the cohorts to fall into distinct phenotypes in the beginning, the movement toward true insights derived from prospective studies is streamlined. Furthermore, this approach simplifies downstream analyses by obviating the need for forced, overly elaborate statistical adjustments to manage confounding in a heterogeneous data set, allowing for clearer interpretation. Cai et al. converge on a similar view, emphasizing the need for tailored recommendations for the overlooked subtypes of patients, such as a persistently active KL3 subgroup compared with a KL2 subgroup. No one can deny the foundational impact of novel initiatives like the OAI. Public data sets promote collaboration, transparency, and reproducibility. As demonstrated by the thoroughness of the study by Cai et al., even advanced joint models that are methodologically airtight are hamstrung from making strong conclusions. We are reaching the limits of what we can understand from this data set. It is time to graduate from the comfort of recycling decades-old data. If we want to realize our goal of personalized, precision medicine, we must look forward and actively build it. It is increasingly difficult to justify using questionnaires from 1990 as the gold standard for insights in 2026, especially when wearable sensors fit into a ring and every patient’s phone is equipped with machine intelligence.
Woo et al. (Wed,) studied this question.