Revisiting the OGIPRO Trial: Dynamic Electronic Patient-Reported Outcomes Compared with EQ-5D-5L in HER2-Positive Breast Cancer.

1. Introduction
Patient-reported outcomes (PROs) have become increasingly important in oncology, providing insights into treatment tolerability and patient well-being that extend beyond traditional clinical endpoints. In parallel, precision oncology and artificial intelligence (AI) approaches increasingly rely on longitudinal, patient-centered data streams to support individualized monitoring and timely clinical decision-making. However, for PRO-derived digital signals to be used confidently in such workflows, they must be anchored to established, clinically interpretable instruments.
The EQ-5D-5L, introduced by the EuroQol Group in 2011, is one of the most widely used standardized instruments for assessing health-related quality of life [1]. It evaluates five core dimensions—mobility, self-care, usual activities, pain/discomfort, and anxiety/depression—each on a five-level scale and includes a visual analogue scale (VAS) of overall health.
Electronic PROs (ePROs), particularly mobile-application-based systems, enable the remote, asynchronous, and structured collection of PROs. Dynamic ePRO describes a class of ePROs, where patients provide data on their own initiative, on their own time, thereby generating higher-frequency longitudinal trajectories than scheduled questionnaires. Prior studies have shown that dynamic ePRO monitoring facilitates the early detection of adverse events, supports timely clinical interventions, enhances patient–physician communication, and shows high adherence [2,3,4,5,6]. ePRO monitoring is feasible even in patients with advanced cancer [7]; digital interventions are increasingly accepted [7] and show the ability of improving symptom management, quality of life and survival [2,8].
Although dynamic ePROs offer advantages in data richness and timeliness, their relationship to established PRO instruments such as the EQ-5D-5L remains insufficiently characterized. Establishing congruence between continuous, patient-initiated ePRO signals and standardized measures is an important prerequisite for integrating such data streams into routine care and for enabling downstream precision-medicine applications, including individualized monitoring and predictive or decision-support models.
We therefore examined the alignment of dynamic ePRO reporting with EQ-5D-5L assessments in patients with HER2-positive breast cancer enrolled in the OGIPRO clinical trial (KEK-ZH 2021-D0051) [9]. Medidux, the dynamic ePRO tool used in the OGIPRO study, allows structured, CTCAE-aligned symptom [10] reporting across more than 110 categories. Using linear mixed-effects modeling, we investigated associations between (i) ePRO well-being and EQ-5D-5L VAS scores; (ii) ePRO symptom grades and EQ-5D-5L domain sums; (iii) ePRO symptom grades and EQ-5D-5L disutility using the EQ-5D-5L value set for Germany [11].

2. Materials and Methods

2.1. Study Design and Population
This analysis was conducted within the OGIPRO trial (KEK-ZH 2021-D0051), a non-interventional, multicenter, prospective observational study in Switzerland. Eligible participants were adults with HER2-positive breast cancer (early-stage, locally advanced, or metastatic) receiving systemic therapy according to national guidelines, see Table 1. Written informed consent was obtained from all patients. This study was approved by the Cantonal Ethics Committee Zurich and conducted in accordance with the Declaration of Helsinki.

2.2. Data Collection
Patients reported symptoms and overall well-being using the Medidux mobile application (version 3.2, mobile Health AG, Zurich, Switzerland), which allows collection of dynamic ePRO data. The app provides structured, CTCAE-based symptom reporting across more than 110 categories and includes a continuous well-being rating. Patients were encouraged to record entries as frequently as they deemed relevant.
In parallel, standardized EQ-5D-5L questionnaires were electronically administered approximately once weekly. Each EQ-5D-5L record included five domain-specific items—mobility, self-care, usual activities, pain/discomfort, and anxiety/depression—each graded on a five-level severity scale, and a visual analogue scale (VAS) reflecting overall health status.
See Figure 1 for an example set of data.

2.3. Data Preprocessing
For comparative analysis, EQ-5D-5L entries were aligned with ePRO reports within a ±1-day window. When multiple ePRO entries fell within this window, averages were calculated separately for well-being and symptom grades. When no ePRO entries were available within this window, the EQ-5D-5L entries were not used in the analysis.
When comparing ePRO symptom grades to EQ-5D-5L domains, it should be noted that EQ-5D-5L domains are usually converted to a health score, called an index, using country-specific weights, called value-sets [1,12,13]. A value-set for Switzerland does not exist, with the German value-set being commonly used as a replacement. The index is calculated using:
Here, U is the resulting index value, D is the disutility, and is the weight value for the EQ-5D-5L domain DOMAIN at a given level level. See Table 2 for the German value-set [11].
Since the ePRO symptom grades, the individual reported EQ-5D-5L levels, as well as the domain sum all increase with worsening health, but the index instead decreases with worsening health, we used only the disutility term, scaled by its hypothetical maximum value (for normalization), yielding D′:
It is debatable whether the average of the ePRO symptom grades are more appropriately reflected by the EQ-5D-5L index or the simple domain sum. We decided to analyze both.
The analytic dataset comprised a total of 3376 well-being and 10,323 symptom grade entries (across 91 different symptom types), for a total of 13,699 ePRO entries from 53 patients. Of these, 50 patients contributed data appropriate for analysis (i.e., both EQ-5D-5L and dynamic ePRO entries within matching time window), see Figure 2.
We averaged the symptom grades across all 91 symptom types into a single average symptom grade. While this reduced the richness of the dataset (see Section 4), performing a more detailed analysis would have required more advanced methods and likely a larger dataset.
After averaging well-being and symptom grades, a total of 252 matched observations from 49 patients were available for the analysis of well-being versus EQ-5D-5L VAS and 226 matched observations from 48 patients for symptom grade versus EQ-5D-5L domain sum/disutility.
All dynamic ePRO and EQ-5D-5L values were normalized to a 0–1 scale. Higher dynamic ePRO symptom grades correspond to greater symptom burden, while higher well-being values represent better well-being: This corresponds to the directionality of EQ-5D-5L, where higher domain grades correspond to greater disutility, but a higher VAS corresponds to better subjective health.

2.4. Statistical Analysis
Associations between dynamic ePRO measures and EQ-5D-5L outcomes were assessed using linear mixed-effects models to quantify both population-level associations and patient-specific deviations in longitudinal PRO trajectories, see Figure 3. We chose a relatively simple over a complex model for interpretability and reproducibility,] and to avoid overfitting and generating a false positive result, in accordance with recent guidance [14].
For the first model, the EQ-5D-5L VAS score was modeled as a function of dynamic ePRO well-being:
where yij and xij are the VAS and well-being entries for patient i at observation j, β is the fixed (population-level) slope, ui is the patient-specific deviation (random effect) from the population slope, and εij is the residual error term.
For the second model, the EQ-5D-5L domain sum was modeled as a function of dynamic ePRO symptom grade, using the same model, with yij and xij being the EQ-5D-5L domain sum and symptom grades for patient i at observation j.
For the third model, we used the EQ-5D-5L disutility (see Equation (2)) instead of the EQ-5D-5L domain sum.
Estimated coefficients β and u, standard errors, and 95% confidence intervals were derived using Restricted Maximum Likelihood (REML) analysis. All analyses were conducted using Python (version 3.13.1, Python Software Foundation, Wilmington, DE, USA) with numpy (version 2.4.2) [15], pandas (version 3.0.0) [16], and the statsmodels package (v 0.14.5) [17]. Plots were created using matplotlib (version 3.10.8) [18] and seaborn (version 0.13.2) [19].

3. Results

3.1. Well-Being vs. EQ-5D-5L VAS Score
A total of 252 observations were analyzed (averages of well-being entries within ±1 day of EQ-5D-5L replies versus EQ-5D-5L VAS score), with minimum, mean, and maximum group sizes (number of matched pairs per patient) of 1, 5.1, and 10, respectively.
Mixed-effects model regression for VAS vs. dynamic ePRO well-being yielded a coefficient β = 1.061 ± 0.024; 95% CI: 1.015–1.107. The variability between groups (i.e., patient-specific deviation from the population slope u) was 0.023 ± 0.070. For visual representation, see Figure 4.

3.2. Symptom Grade vs. EQ-5D-5L Domain Sum
A total of 226 observations were analyzed (averages of symptom grade entries within ±1 day around EQ-5D-5L replies versus EQ-5D-5L domain sum), with minimum, mean, and maximum group sizes (number of matched pairs per patient) of 1, 4.7, and 10, respectively.
Mixed-effects model regression for EQ-5D-5L domain sum vs. dynamic ePRO symptom grade yielded a coefficient β = 0.404 ± 0.049; 95% CI: 0.307–0.501. The variability between groups (i.e., patients-specific deviation from the population slope u) was 0.093 ± 0.394. For visual representation, see Figure 5.

3.3. Symptom Grade vs. EQ-5D-5L Disutility
A total of 226 observations were analyzed (averages of symptom grade entries within ±1 day around EQ-5D-5L replies versus EQ-5D-5L disutility), with minimum, mean, and maximum group sizes (number of matched pairs per patient) of 1, 4.7 and 10, respectively.
Mixed-effects model regression for EQ-5D-5L domain sum vs. dynamic ePRO symptom grade yielded a coefficient β = 0.213 ± 0.032; 95% CI: 0.151–0.275. The variability between groups (i.e., patient-specific deviation from the population slope u) was 0.041 ± 0.322. For visual representation, see Figure 6.
For a tabular view of the model results for all 3 models, see Table 3.

4. Discussion
This study provides the first real-world comparison of continuous ePRO monitoring with EQ-5D-5L questionnaires in patients with HER2-positive breast cancer receiving systemic therapy. By generating high-frequency, patient-initiated longitudinal data, the ePRO system captured a richer and denser dataset than scheduled questionnaires, underscoring its potential for supporting individualized patient monitoring and data-driven clinical workflows.
Our findings demonstrate that self-reported well-being determined via Medidux aligns closely with EQ-5D-5L VAS scores, with nearly one-to-one correspondence and minimal between-patient variability. This indicates that dynamic ePRO well-being represents a robust and clinically interpretable digital phenotype that can be reliably used as a substitute for EQ-5D-5L VAS collection in both clinical practice and research, providing a validated high-resolution input for continuous patient monitoring and downstream analytical applications.
By contrast, the relationship between the aggregated ePRO symptom grades and EQ-5D-5L domains was weaker and more variable. This held true regardless of using the EQ-5D-5L domain sum or country-corrected index. This heterogeneity likely reflects the loss of information introduced by aggregation: Medidux captures detailed, multi-dimensional symptom profiles across many categories, whereas the EQ-5D-5L compresses patient experiences into five broad domains, and our analysis further reduced both to single summary measures. Collapsing such complex, high-dimensional patient-reported data into global averages may obscure clinically relevant, patient-specific symptom patterns. It is plausible that specific symptoms correlate more strongly with some EQ-5D-5L domains than with others, e.g., the dynamic ePRO fatigue symptom probably more strongly correlates with the anxiety/depression or the mobility domain than with the pain/discomfort domain. Since 91 symptom types × 5 EQ-5D-5L domains yielded 455 potential correlations, we decided against analyzing them with mixed-effects models. This rich data lends itself to more advanced analytics, which should be explored in future research.
Another important limitation is the fact that we did not control for additional factors such as age, performance status, cancer stage, comorbidities, or other individual characteristics that may also contribute to the observed variability between patients and where there may be group differences. Given the small sample size, we consciously decided against it.
Lastly, the population analyzed consist entirely of women from Switzerland with HER2-positive breast cancer. This limits the generalizability of the results.
Future studies should examine symptom–domain relationships with finer granularity and in larger cohorts, enabling more detailed mapping between specific symptom clusters and quality-of-life domains. In this context, machine learning and other advanced AI methods may be particularly well-suited for modeling the high-dimensional, longitudinal structure of dynamic ePRO data, identifying clinically meaningful patterns, and capturing patient-specific trajectories that cannot be resolved through global summary scores alone. Such approaches may be required for fully leveraging continuous ePRO data streams in precision oncology, where individualized symptom dynamics can inform personalized monitoring strategies and provide data-driven clinical decision support. We believe that integrating dynamic ePROs with other digital biomarkers, wearable device data, and electronic health records is a promising area of future research.

5. Conclusions
Dynamic ePRO reporting via Medidux shows strong concordance with EQ-5D-5L well-being measures while providing higher temporal resolution and richer clinical context. These results support the use of dynamic ePRO well-being data as a pragmatic replacement for EQ-5D-5L VAS in both clinical practice and research, enabling continuous, patient-centered monitoring. For domain-level outcomes, further methodological refinement and larger datasets are required before dynamic ePRO symptom grades can be substituted for EQ-5D-5L domains.
Beyond questionnaire replacement, dynamic ePRO systems represent a scalable source of longitudinal, patient-generated data that is well-suited for precision oncology applications. By providing validated, high-frequency digital phenotypes, such systems can support advanced analytic approaches, including predictive modeling and AI-driven decision support, with the potential to improve individualized patient management and clinical outcomes.