data
The data used in this study was collected in the Pro B Research 16. The PRO B study was a random contrast health service research test of multi -interceptors between May 2021 and February 2024, which registered 924 patients hired by 52 medical centers across Germany. Eligible patients were 18 years of age or older who were able to read and understand German, and have been treated for anticancer drugs for metastatic breast cancer, and have been expected to receive more than three months of life. Additional qualifications include Internet access through mobile phones and ECOGs (ECOGs) and ECOG performance status of 0 to 2. The patient is stratified and randomly assigned to a 1: 1 ratio based on research centers, transition sites, and undercombine types. Then, when the Pro Value worsened, it was assigned to the control group 16, which completed the Pro irradiation every three months without the application of an application -based daytime Pro monitoring (followed by doctor contact and individual treatment) or a warning. The patient of the arbitration group completed a different short form every week at the EORTC CAT Bank 17, while the patient of the control group completed every three months. The EQ-5D-5L was evaluated at the baseline, 6 months and 12 months through the mobile phone application of two groups. The German price set was used to calculate the utility index 18. Observations with completely missing data for EOTC items or EQ-5D-5L were excluded from the analysis. The PRO B research design did not skip the PRO item, so no data missed from the item level was not observed. Therefore, omission data is based on complete and reliable data without confrontation.
data set
The initial data set contains 2,474 observations from 909 patients (15 patients were excluded from PRO-B studies because they did not provide a complete response to the questionnaire). After 635 observations with missing data for the EORTC item or EQ-5D-5L, the data set contains 1,839 observations from 878 patients (Figure 1). Data from the EOTC and EQ-5D-5L were fully fully in the baseline, 6 months and 12 months, and clustering of multiple reactions per patient was adjusted in the mapping model. Mapping models often work well in the same data as used to develop, so you need to verify the external data set to evaluate the actual mapping performance. But in our study, external data sets cannot be used. Therefore, we used an internal verification sample in this study divided into two sets through the computer creation random number. 70% of patients (1,269 observations, 609 patients) were randomly assigned. Estimated set The remaining 30% (570 observations, 269 patients) to create a mapping model Validation set Test model performance. In order to ensure a similar disease severity distribution between estimation and verification sets, the EQ-5D-5L Utility Index for all patients has been grouped into four initiatives (0.775, 0.775–0.876, 0.877–0.942, ≥ 0.943) over all time. For example, if patient A has a real utility index value of 0.760, 0.913, and 0.861, respectively, this value is assigned to the utility group 1, 3 and 2, respectively, if the actual utility index value of 0.760, 0.913 and 0.861, respectively. Then, based on a specific pattern of utility grouping, the patient is stratified (eg 1, 2, 3 or 3, 2, 1, and equal), ensures that the group’s sequence was preserved while distinguished from other groups such as 1, 2 or 4. In accordance with this pattern group, a randomized patient was assumed or validated in a state in which a randomized patient was assumed or validated. It was distinguished from the patient, valid and verified.
musical instrument
We conducted the EQ-5D-5L evaluation through the German version of the EQ-5D-5L questionnaire 15, which includes five items: mobility, self-management, general activity, pain/discomfort and anxiety/depression. In the EQ-5D-5L, the patient evaluates the health status from 1 (no problem) to 5 (extreme problems) for each of these five dimensions. We converted the EQ-5D-5L response to the utility index through the German Value Set 18 to describe the consideration of social and national considerations for the HRQOL difference, converting the reaction to a single index that indicates the overall health status. The German value set provides a utility index from -0.661 to 1, where 0 shows the same health as death, and the value of less than 1 can be considered worse than dead.
QLQ-C30 (version 3.0) is a standardized cancer questionnaire developed by EORTC to evaluate HRQOL in cancer patients. It includes 30 items including functional health, symptoms and full QOL (19). In this study, HRQOL was evaluated in several short forms of EORTC Cat Core item Banks (17), which was specifically developed for PRO B research. This short form was used as a source measurement of the mapping algorithm and consisted of 51-73 items per domain. The QLQ-C30 consists of 30 items that assess nine multiple items, but all functional health and symptom areas of QLQ-C30 are included in the EORTC Cat Core item Bank. EORTC CAT CORE and QLQ-C30 includes five functional scales (physical, role, cognition, emotional and social function), nine symptoms (fatigue, pain, nausea, vomiting, insomnia, greed, constipation, diarrhea and financial difficulties). EORTC CAT core scores use standardized use tea A score based on the normative indicators of a general population with an average of 50 points and a standard deviation of 10 points (20). Meanwhile, the QLQ-C30 score is calculated by average the items within each scale and then converted to the range of 0-100 (21). The high value of the functional scale shows a good function, while the symptom scale or the high value of the item shows a high symptom burden.
Statistical analysis
All statistical analysis was performed through STATA 18/MP (College Station, Texas, USA). The baseline properties are presented through the description statistics including the proper average, standard deviation (SD), minimum, maximum and percentage for estimated and verified sets. By repeated measurement correlation coefficient, we explored the relationship between the EORTC CAT core, QLQ-C30 and EQ-5D-5L.RM), Explaining the volatility between objects implemented in R package)RMCORR”(22).
We followed the ISOR’s best case guidelines for mapping general health status in the non -reference -based device 23, and reported the results according to the method of mapping for the standard checklist 24 (supplementary table S1 and S2).
For all direct and indirect mapping models, we performed the entire model that includes all the domains of EORTC as an independent variable. For direct mapping, the dependent variable was EQ-5D-5L Utility Index, and based on previous studies, four regression types of OLS, TOBIT, adjusted beta regression and ALDVMM were used. According to the previous EQ-5D mapping study, the OLS model assumes that the EQ-5D-5L utility index may be presumed to be a linear combination of the response of the EORTC questionnaire. Some studies have shown that OLS is best performed when the QLQ-C30 is mapped to EQ-5D index values (3, 10, 13, 14). However, this model can produce an inappropriate estimated value due to the ceiling effect of the EQ-5D index 25. To solve this boundary utility problem, both the Tobit model and ALDVMM were applied to set the lower and upper limits to -0.661 and 1, respectively. ALDVMM expresses the flexibility of modeling, boundary and multimodal distribution of health utility by expressing health utility in a mix of continuous distribution (26). Each observation is likely to contribute to each component, which can be flexibly approximating the complex utility distribution. In order to identify the global maximum value of the urethral function in ALDVMM, we have improved the possibility of convergence in the world by using both local and global optimization approaches. First, we installed a constant model and initialized the entire model using “” “using the estimated parameters.Inimethod“Options ALDVMM (23) Command. In addition, we used global optimization technology using simulation annealing with different start values and arbitrary seeds to explore the parameter space extensively. Then we compared the results of the two methods using Log-Likelihood and Bayesian Information Criterion (BIC). In our study, local optimization suggests that local optimization has been effectively identified for global solutions with higher possibilities and better model suitable (BIC -based) than global optimization. We have searched for models with up to five components. However, after adjusting the optimization setting and increasing the repetition, four and five ingredients were not converged. Based on convergence, analysis and BIC, we maintained 1, 2 and 3 component models for further analysis. In the case of the adjusted beta regression requiring scores between 0 and 1, the problem of the EQ-5D-5L Utility index was converted to 0-1 scale to solve the problem of negative water. This conversion was carried out through the following formula. (Observed Utility- (-0.661))/(1 -(-0.661)), where -0.661 represents the lowest utility index in the German value set (18). The predicted utility index was well -equipped with one if it was higher than 1.
For indirect mapping, we have performed a generalized order Loit model for each EQ-5D-5L item (Seosu dependent variable) through the STATA command.GOLOGIT2(27). After that, the estimated response was combined, and the utility value was calculated through the German value set 18. Since each item was modeled separately, each mapping algorithm consists of five separate models. It was used to estimate the probability, but its specificity could be a specific approach to analyze the proportional probability assumptions (27, 28).
Model verification and prediction ability
We report the measurement of the entire model suitable, including the average absolute error (MAE), the root average square error (RMSE), the average prediction buyer (observed average -predicted average), and the matching correlation coefficient (CCC) of the LIN. In addition, the density distribution function is presented to compare the actual and estimated data to evaluate the model suitability across the distribution of possible values. To assess the potential systematic bias between the actual utility and the estimated utility, the Bland Altman Plots 29 shows a difference in the average of the utility index value observed and the average of this value. In addition, we included Lowess smoothing as 95% trust section. In order to assess the consistency of the mapping algorithm in the range of various utility indexes, the prediction accuracy was investigated compared to the total range of EQ-5D-5L Utility Index values and focused on the intervals of 0.60, 0.60–0.79 and 0.80-1.00.
