Prospective multicenter assessment of patient preferences for properties of gadolinium-based contrast media and their potential socioeconomic impact in a screening breast MRI setting

We conducted a prospective observational discrete choice conjoint survey at 4 institutions in Michigan, Minnesota, New York, and Indiana from July 2018 to March 2020. The institutions were chosen to support a broad range of demographic characteristics. Patients participating in breast MRI screening programs are known to have a skewed demographic distribution compared to the general population [9‐13]. The inclusion criteria were chosen to reflect individuals with a vested personal interest in cancer detection and a high likelihood of repeated lifetime administrations of GBCM. Inclusion criteria were as follows: (1) outpatient, (2) intermediate or high risk for breast cancer [14] undergoing annual screening breast MRI. Exclusion criterion was previous participation in the study (N = 0). Indications for MRI were based on the American Cancer Society’s 2007 guidelines for breast cancer screening with MRI [14].

Our study design used a paired discrete choice-based conjoint survey. This type of survey provides respondents with 2 options and asks them to select which one they would prefer. Each option has a series of attributes that the investigators wish to study. In our context, those attributes were the risks and benefits of GBCM. That process is repeated multiple times. At the conclusion of the survey, implied preferences can be derived that indicate what attributes were prioritized by the respondents when making their choices.

Our discrete choice-based conjoint survey (Supplemental Material; Sawtooth Software, Inc (Provo, UT, USA); [15]) used a partial profile design and provided respondents with 15 paired choice sets. Each choice was between two unique hypothetical GBCM with the same 5 attributes set at the same or different levels (sensitivity for cancer detection [range 80–95%], intracranial gadolinium retention [range 1–100 molecules retained per 100 million molecules administered], severe allergic-like reaction rate [range 1–19 per 100,000 administrations], mild allergic-like reaction rate [range 10–1000 per 100,000 administrations], and out-of-pocket cost [range $25–$100]). Only one answer was allowed per question. The range of most GBCM attribute levels was derived from the literature (cancer detection sensitivity [16, 17], gadolinium retention [18], severe allergic-like reaction rate [19], and mild allergic-like reaction rate) [19]. Out-of-pocket cost was informed by common co-pay rates, and online drug prices and proprietary vendor-negotiated price contracts from participating institutions. The content of the survey was vetted by patient advocates with experience in survey design and underwent precognitive pilot-testing for content and readability by five patients undergoing breast MRI who were not included in the study and who met inclusion criteria. A professional medical illustrator created infographic information to facilitate patient understanding.

The survey software considered the active ratings of the respondents to generate a personalized choice set that maximized analyzability of their responses. The experiment had a near-orthogonal design with level balance and minimal attribute level overlap. Details of the administered surveys are provided in Supplementary Table 1.

The survey was administered by trained interviewers. Potential participants were recruited by reviewing the daily breast MRI schedule at each participating institution and consenting patients in real-time the day of their examination. Demographic data (patient age, patient ethnicity, indication for MRI) were extracted from the electronic medical record for all screened patients at each site to (1) reduce the question burden from each patient, and (2) to characterize the non-respondent population.

An a priori power calculation was performed to estimate the needed sample size. Based on the largest observed standard deviation available from preliminary data, to achieve a ± utility confidence interval length of 10 required 170 patients, and to achieve a ± utility confidence interval length of 5 required 670 patients. The study was terminated after 236 patients had been accrued due to the emerging novel coronavirus disease 2019 (COVID-19) pandemic [20].

Hierarchical Bayesian modeling and a Monte Carlo Markov chain algorithm were used to estimate part-worth utilities (and their 95% confidence intervals) for each GBCM attribute [21‐23]. A total of 50,000 posterior simulation iterations were used. Part-worth utilities were an interval measure of patient preference for levels within an attribute—somewhat analogous to a beta coefficient from a logistic regression. To aid interpretation, part-worth utilities were zero-centered so that positive values indicated increased likelihood of selection and negative values indicated decreased likelihood of selection.

Attribute importance is the estimated average relative importance participants placed on a given attribute when making GBCM selection decisions. For each participant, attribute importance was calculated as the range of their part-worth utilities for that attribute, divided by the sum of the ranges for all attributes, multiplied by 100 (i.e., \( \frac{Specific\ attribute\ utility\ range}{\sum All\ attribute\ utility\ ranges}\times 100 \)). The average attribute importance was reported as a percentage and a 95% confidence interval. Attribute importances summed to 100%.

Demographic data were summarized with descriptive statistics. Patient-specific attribute importances were modeled using linear regression. The covariates used for analysis included age, education, health insurance, employment, household income, and previous allergic-like reaction to GBCM. The mean attribute importance difference was calculated between each covariate subgroup.

Statistical analysis was performed using SAS software (v9.4). For primary endpoints, p < 0.05 is considered significant. For secondary endpoints (i.e., when assessing differences in attribute importance), p < 0.01 was considered significant to account for multiple comparisons.

We created a simulator that could be used to ascertain how current and future (i.e., hypothetical) GBCM products would perform in the marketplace (relative to each other) if GBCM selection was solely based on the patient preference data we collected in our multi-site study. Utility values for a combination of GBCM properties were combined to build a multi-product competitive model using the randomized first choice method to estimate share of preference. Patient-specific share of preference for a GBCM was calculated as the antilog of the total product utility (based on patient-specific part-worth utilities). Results for each product were rescaled to sum to 100%. Overall share of preference was calculated as the mean of patient-specific shares of preference. Patient-centered simulations were performed to compare 3 existing GBCM (using published data) with 3 hypothetical GBCM using the patient-level part-worth utilities derived from our study. The specific GBCM property information is included in Table 1. Although these six GBCM (3 existing, 3 hypothetical) were the only GBCM formally analyzed in our simulations, the simulator derived from our data (Supplementary Material) permits the user to input any combination of attribute levels based on existing or novel GBCM to determine its hypothetical value (from the patient’s perspective) vs. a user-defined number of competitor GBCM.

The survey response (87% [247 of 284]) and completion (96% [236 of 247]) rates were excellent. A total of 284 patients were approached to participate; 247 agreed to participate and 236 completed the survey. Incomplete surveys (N = 11) were excluded from analysis. A study population flow diagram is provided in Fig. 1. Non-responders (N = 37) had similar demographics to responders (N = 236) (Table 2). Recruitment distribution from the four sites was as follows: site 1 (38% [90 of 236]); site 2 (30% [70 of 236]); site 3 (16% [38 of 236]); site 4 (16% [38 of 236]). Details of our study population are provided in Table 2.

A majority (85% [201 of 236]) of participants were white with a mean age of 50 ± 11.9 years. Most had a household income greater than $75,000 (77% [158 of 204] among those reporting income), were college educated (70% [164 of 236]), had full-time employment (64% [150 of 236]), and had an employer-based insurance plan (75% [176 of 236]). Only 2% (5 of 236) reported a previous allergic-like reaction to GBCM.

The values respondents ascribed to each GBCM attribute, expressed as part-worth utilities, are included in Fig. 2, with pairwise comparisons between levels provided in Supplementary Table 2. Participants preferred (p < 0.001) greater cancer detection sensitivity, lower cost, less intracranial gadolinium retention, and lower mild and severe allergic-like reaction rates (Fig. 2, Supplementary Table 2). Patients considered cancer detection sensitivity to be the most important GBCM attribute (Fig. 2, Table 3). GBCM attributes are listed as follows in descending order of importance (%, Table 3): (1) cancer detection sensitivity (44.3%, 95%CI 42.0–46.7%), (2) mild allergic-like reaction rate (19.5%, 95%CI 17.9–21.1%), (3) severe allergic-like reaction rate (17.0%, 95%CI 15.8–18.1%), (4) intracranial gadolinium retention (11.6%, 95%CI 10.5–12.7%), (5) out-of-pocket cost (7.5%, 95%CI 6.8–8.3%). Intracranial gadolinium retention was considered by patients to be less important than cancer detection sensitivity and allergic-like reaction rates (mild and severe), but more important than out-of-pocket cost.

Attribute importance varied by several factors including household income and educational attainment (Table 3, Supplementary Table 3). Compared to patients with an annual income > $150,000, patients with an annual income < $25,000 placed less value on cancer detection sensitivity (importance 33.1% vs. 48.8%, p < 0.01), greater value on out-of-pocket cost (importance 14.9% vs. 5.7%, p < 0.01), and greater value on intracranial gadolinium retention (15.7% vs. 10.1%, p < 0.01) (Table 2). College-educated patients placed slightly less importance on out-of-pocket cost (−2.8%, 99%CI −4.9 to −0.7) (Table 3). However, the rank-ordering of attribute importance between these income and educational strata was nearly the same, and cancer detection sensitivity always had the highest importance. Age, insurance status, employment status, and level of breast cancer risk did not have a significant effect on GBCM attribute importance (Table 3). Race and ethnicity subgroups were insufficiently powered for subgroup analysis.

Multi-product competitive simulations were performed to determine which GBCM a population of patients would prefer if they knew the details and were empowered to choose which GBCM they were administered (Table 4). In all simulations, existing product C (macrocyclic, high sensitivity 94%, low intracranial gadolinium retention 0.2 per 100 million molecules, high mild reaction rate 150/100,000, low severe reaction rate 5.7/100,000, intermediate out-of-pocket cost $72) was the most preferred (preference share 57.4–61.7% competing vs. 2–5 other GBCM) (Table 4). This was driven by superior sensitivity for cancer detection, low gadolinium retention, and low severe reaction rate (Table 3).

Hypothetical macrocyclic GBCM (products E and F) with low cancer detection sensitivity (78–83%), low intracranial gadolinium retention (0.1–0.2 per million molecules), and intermediate to high allergic-like reaction rates (mild, 72–130/100,000; severe, 12–18/100,000) were less preferred than two existing linear GBCM (products A and D) with similar cancer detection sensitivity (83%), intermediate to high gadolinium retention (4.5–20 per million molecules), and low allergic-like reaction rates (mild, 12–39/100,000; severe, 1.6–2.1/100,000) (Table 4). Even though the two hypothetical macrocyclic agents had lower intracranial gadolinium retention, the higher patient importance for low mild and severe reaction rates made the linear agents A and D (preference shares: 8.0% and 12.0%) more preferred than macrocyclic agents E and F (preference shares: 1.6% and 0.9%) (Tables 3 and 4).