Validation of Four EQ-5D-5L Crosswalk Prediction Models From Promis-29 in Patients With Cardiovascular Disease
Author(s)
Klapproth C1, Fischer F2, Liegl G3, Martin CN4, Blumrich A4, Rönnefarth M4, Schmidt S4, Rose M5
1Charité – Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, BE, Germany, 2Charité – Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, Germany, 3Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, BE, Germany, 4Berlin Institute of Health at Charité (BIH), Berlin, Germany, 5Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, Germany
Presentation Documents
OBJECTIVES: The EQ-5D-5L crosswalk (EQ-5D) is a preference-based score to estimate quality-adjusted life years (QALY) in cost-effectiveness analyses. The descriptive PROMIS-29 profile is a patient-reported outcome measure used in clinical routine and research. Four different mapping models are available to predict the EQ-5D from PROMIS-29 scores, but these have not yet been tested in independent data. We sought to investigate their prediction performance, in particular when assessing change over time.
METHODS: We analyzed data from the Berlin Longterm Observation of Vascular Events Study (BeLOVE), an observational clinical cohort study of patients with cardiovascular disease. In the baseline sample of n1=1118, we assessed prediction performance as EQ-5D1 – ÊQ-5D1. In the follow-up sample of n2=565, we assessed prediction performance of changes in EQ-5D as (EQ-5D2 – EQ-5D1) – (ÊQ-5D2 – ÊQ-5D1). As measures of accuracy and agreement, respectively, we used the normalized root mean square error (nRMSE), the normalized mean absolute error (nMAE), intraclass correlation coefficient (ICC), R2, and Bland-Altman plots.
RESULTS: Across all prediction models, we observed little to substantial bias (-0.10-0.01) between observed and predicted EQ-5D scores. They performed similarly in terms of nRMSE (0.09-0.12), nMAE (0.06-0.08), R2 (0.66-0.73), and ICC (0.77-0.83). Bland-Altman plots showed that one model overestimated the EQ-5D substantially in poorer health states. When predicting change in EQ-5D over time, the ICC (0.42-0.49) was considerably lower. Health improvements were underestimated and health deteriorations were overestimated by an average of about +/-0.15, at most. The greater the improvement (deterioration), the greater the overestimation (underestimation).
CONCLUSIONS: We found that by most of our applied measures, all four models perform similarly well. However, one model showed substantial bias. One other model performed worse in predicting poorer health states, probably because it is not polynomial as the others are. In addition, agreement between observed and predicted longitudinal changes in EQ-5D is only fair.
Conference/Value in Health Info
Value in Health, Volume 25, Issue 12S (December 2022)
Code
MSR12
Topic
Methodological & Statistical Research, Patient-Centered Research, Study Approaches
Topic Subcategory
Decision Modeling & Simulation, Health State Utilities, PRO & Related Methods
Disease
No Additional Disease & Conditions/Specialized Treatment Areas