Do Personalized Recovery Trajectories Enhance the Utility of Patient-reported Outcome Measures After Orthopaedic Surgery?

Patient-reported outcome measures (PROMs) could be useful tools for supporting patient care after orthopaedic surgery, but they can be difficult to interpret and apply in practice. Existing methods for PROM interpretation, such as meaningful change values and population norms, have practical limitations and are unresponsive to patient-specific factors that influence recovery. Personalized recovery trajectories could help overcome these limitations, making PROM scores more clinically meaningful and actionable. (1) Does a "people-like-me" modeling approach to crafting personalized recovery trajectories provide accurate and precise trajectories of PROM recovery early after total shoulder arthroplasty (TSA)? (2) What advantages do personalized recovery trajectories provide over alternative methods for PROM interpretation (minimum detectable change, minimum clinically important difference, and normative values) when applied to individual patients? We used data from 28 outpatient physical therapy clinics within a single integrated health system from 2017 to 2022. We included patients who underwent unilateral anatomic or reverse TSA and had postoperative data available from the DASH questionnaire. We excluded patients who were missing key modeling variables, who had fewer than three postoperative DASH assessments, or who lacked a DASH assessment within the first 45 days after surgery to ensure adequate data for modeling early recovery. Based on those criteria, we included 60% (1392 of 2322) of eligible patients in our primary analysis. The median (IQR) patient age was 68 years (62 to 74), and 51% (710 of 1392) were women. Excluded patients had slightly lower median BMI (29.0 versus 30.0 kg/m2), were more likely to undergo anatomic TSA (64% [591 of 930] versus 59% [816 of 1392]), were less likely to have commercial insurance (28% [258 of 930] versus 35% [483 of 1392]), and were more likely to have Medicare (65% [608 of 930] versus 62% [859 of 1392]), Medicaid (3% [32 of 930] versus 2% [26 of 1392]), or workers' compensation (3% [30 of 930] versus 2% [24 of 1392]) insurance. However, these differences were small and unlikely to substantially impact our study's generalizability. To answer our first question, we split eligible patients into training (n = 928) and testing (n = 464) groups to develop a "people-like-me" model that predicts DASH scores after TSA based on the recovery of historical patients with similar characteristics. We then used the "people-like-me" model to generate personalized trajectories of DASH recovery over the first 120 postoperative days for patients in the testing group. We modeled the acute recovery period because PROMs are most clinically actionable early after surgery, and later time points had little data available. To answer our second question, we included all patients from the testing group who had preoperative DASH scores available (n = 20) to illustrate the potential advantages of using personalized recovery trajectories for DASH interpretation. We then selected three of these patients for case studies to compare personalized recovery trajectories against minimal detectable change (9 points) and minimum clinically important difference (12 points) values from previous research and against normative values derived from the training group. The DASH is scored from 0 to 100, where higher scores indicate greater disability. The model demonstrated good accuracy; the mean error (mean of the differences between predicted and observed DASH scores) was -2 points, and the mean absolute error (mean of the absolute differences between predicted and observed DASH scores) was 11 points. In other words, predicted DASH scores were not systematically overestimated or underestimated, and they were a mean of 11 of 100 points away from patients' true scores. These error rates were relatively small given the considerable variability of DASH scores in the testing group. The model also provided accurate and precise estimates of uncertainty based on the 50%, 80%, and 90% prediction intervals. The proportion of observed DASH scores that fell within each prediction interval was close to the target rate. For example, 49% of observed DASH scores fell within their 50% prediction interval (target rate = 50%). The model's precision was quantified by the mean width of the 50%, 80%, and 90% prediction intervals, which was 18, 34, and 44 points, respectively. In other words, the model expects that a patient's true DASH score will be within ± 9 points from their predicted score 50% of the time, within ± 17 points 80% of the time, and within ± 22 points 90% of the time. These intervals provided a much narrower range of expected DASH scores compared with the entire range of observed scores in the testing group. In the case study comparisons, we showed how personalized recovery trajectories provide context beyond minimal detectable change, minimum clinically important difference, and normative values. For example, personalized recovery trajectories can be applied at any point during the acute recovery period to show whether a patient's DASH score is on track relative to their baseline prognosis, and they account for patient-specific factors that influence postoperative recovery such as age, gender, and BMI. Clinicians should consider using personalized recovery trajectories to make PROMs more useful in the early postoperative period for monitoring patient recovery, supporting interdisciplinary communication, and managing patient expectations. Clinicians can implement personalized recovery trajectories into their practice using tools like the one developed in this study (available at: https://cu-restore.shinyapps.io/TSA_recovery_v1/). Future research should focus on developing and validating similar tools in new populations, integrating them within clinical workflows, and evaluating their impact on patient outcomes. Level III, prognostic study.