Elfi Verheul

Quality of care is a central yet complex concept within healthcare. According to the widely adopted framework by Donabedian, quality can be conceptualized across three core domains: structure, process, and outcomes. These domains form the foundation for developing quality indicators that enable healthcare to be systematically measured and compared. While all three domains are important, modern health systems increasingly prioritize outcomes. These outcomes include both clinical endpoints, such as complications and survival, and patient-reported outcomes (PROs) that measure experienced quality of life.

The systematic measurement and application of outcome information provide opportunities to improve healthcare systems and to inform individual treatment decisions. At the same time, this requires high standards for the validity, reliability, and interpretation of outcome information. Differences in patient populations between hospitals (case mix), statistical uncertainty (reliability), and data quality can all influence results. A rigorous methodological approach is therefore essential to apply outcome information in a mean ingful way.

This thesis investigates how outcome information can be utilized to ensure both reliability and meaningful interpretation. Part I focuses on the use of valid and reliable outcome information to stimulate continuous quality improvement through hospital comparisons, whereas Part II demonstrates how prediction models can translate outcome information into personalised decision-making and expectation management in clinical practice.

Part I: Using outcome information to stimulate continuous quality improvement
Chapter 2 presents a literature review that assesses which cancer outcomes are used for hospital comparisons (i.e. benchmarking) in Europe and how case-mix adjustment is applied. The study shows that benchmarking is conducted in only a few countries, often focusing on short-term outcomes (≤90 days) and lacking standardized or validated case-mix models. Moreover, there is limited transparency regarding the applied case-mix methodologies, which restricts comparability both within and across countries. These findings highlight that benchmarking of cancer outcomes is currently fragmented and methodologically underdeveloped. There is an urgent need for harmonized definitions, standardized approaches to case-mix modelling, and transparent reporting to enable outcome-based care to be compared and improved at national and international levels.

Chapter 3 addresses the validity of quality indicators and describes the development of a case-mix model that adjusts for differences in patient and tumour characteristics be tween hospitals. Such adjustment helps to ensures that observed differences in hospital performance reflect true differences in quality of care rather than variations in patient case mix. This chapter specifically focuses on the quality indicator complications after breast cancer surgery. The results show that, after stratification by type of surgery, ad ditional case-mix adjustment had negligible effects on hospital differences. The patient and tumour characteristics determining complication risk were evenly distributed across hospitals. This suggests that, for this indicator, stratification alone is sufficient and further adjustment is unnecessary. However, such evaluations should be performed for each indicator individually, particularly for outcome indicators where case-mix adjustment typically has a greater impact.

Chapter 4 introduces a framework to systematically assess quality indicators based on four measurable criteria: feasibility (data availability), discriminatory ability, validity (case mix), and reliability. This framework was applied to the Dutch Breast Cancer Audit (NBCA: NABON Breast Cancer Audit). Most NBCA indicators demonstrated high data availability and moderate to good discriminatory ability. For a few indicators, case-mix adjustment is necessary. However, reliability was often a concern, as many observed hospital differ ences were largely driven by random variation, requiring caution in public reporting. his chapter demonstrates that many existing quality indicators are well suited for internal quality improvement but not yet for external public comparison. Reliability should therefore play a central role in decisions about public disclosure. Furthermore, combining data across multiple years substantially improves reliability, offering a practical solution. The proposed framework thus provides a concrete tool to systematically evaluate quality indicators and enhance the quality and transparency of indicator sets.

Chapter 5 applies the framework from chapter 4 to a low-volume disease (oral cavity cancer). Sample size calculations were performed to determine how many patients, and consequently how many years of data, are required for reliable reporting. This provides practical guidance on the minimum reporting period needed to detect quality differ ences with sufficient reliability.

Part II: Using outcome predictions to support personalised care
While the first half of this thesis uses outcome information to understand differences between hospitals, the second half applies the same type of information to personalise care for individual patients. We focus on prediction models that can either support shared decision-making (Chapters 6 and 7) or help manage patient expectations (Chapter 8). Chapter 6 externally validates the latest version of the prediction model PREDICT Breast (v3.1) in Dutch and Swedish cohorts, with particular attention to young patients (≤40 years) and those with lobular breast cancer. External validation means evaluating an existing model in a different population than the one in which it was developed, to assess whether it performs accurately and reliably in new settings. PREDICT Breast v3.1 performed well in the general population and in patients with lobular breast cancer but was less accurate for younger patients. These findings show that PREDICT can be broadly applied in clinical practice, but that model performance is not uniform across all sub groups. This underscores the importance of ongoing validation and subgroup-specific recalibration to ensure the reliability of personalised survival predictions.

Chapter 7 compares the performance of PREDICT Breast v3.1 with that of the previous version (v2.2) in the Dutch population. Both models demonstrated accurate predictive performance, with only minor differences between subgroups. These findings, together with those from Chapter 6, show that external validation is an efficient and valuable strategy: it strengthens clinical confidence and prevents the fragmentation that arises from continuously developing new models from scratch. Systematic validation and reca libration of existing (and, where possible, already implemented) models is therefore often more effective and sustainable than developing multiple overlapping models.

Chapter 8 develops and externally validates fifteen prediction models for various domains of health-related quality of life (HRQoL) after breast cancer surgery. In this case, new model development was justified, as no previous models existed for HRQoL outcomes and a novel methodological approach was applied that analyses trends over time. The preoperative PRO score (baseline measurement) of a specific domain was the strongest predictor of the postoperative trajectory in that domain. The models performed acceptable in external validation and offer potential for use in personalised expectation management. The reason these models are not suitable for shared decision-making about treatment choice is that they were developed using real-world data, in which patients were not randomised to a treatment. As a result, there is selection bias: the patient groups undergoing each type of surgery are not comparable. Therefore, the model can only be applied once the choice of surgical procedure has been made. Nevertheless, this chapter represents an important step towards the use of patient-reported outcomes in prediction care, where managing patient expectations plays a crucial role. However, further refine ment and validation are needed before integration into electronic care pathways can be considered.

Conclusion
This thesis highlights both the potential and the challenges of using outcome informa tion to improve the quality of cancer care. It emphasizes the importance of robust meth odological approaches and transparent reporting to enable valid comparisons of hospital performance. The reliability of quality indicators remains an important consideration, and combining outcome data across multiple years provides a practical way to strengthen this reliability.

Furthermore, validity is crucial for prediction models: external validation ensures the accuracy and generalizability of models across different patient groups and healthcare settings. This should take priority over developing new models, as it is more efficient and, when performance is good (as with PREDICT Breast), it strengthens confidence in their clinical application. Finally, predicting health-related quality of life (HRQoL) represents an important yet challenging next step toward further personalising cancer care.

Recommendations
1. Healthcare systems and registries should adopt a methodological framework for developing and validating quality indicators, with explicit attention to validity, reliability, and transparent reporting of methodology.
2. Public reporting of hospital-level quality indicators should be based on multi-year data to minimize random variation (statistical uncertainty) and improve reliability.
3. Priority should be given to the external validation of existing prediction models over the continuous development of new ones.
4. Prediction models for health-related quality of life (HRQoL) offer a new opportunity to further personalise care. Continued refinement, validation, and integration into clinical practice are needed to fully realize this potential.