Background and relevance
Over 60% of cancer patients receive radiotherapy. Recently, new radiotherapy techniques have been introduced, including proton therapy and MRLinac, that aim to provide the same tumour control, but with reduced irradiation to the surrounding healthy tissues, thereby reducing radiation-induced toxicity. Traditionally, randomized controlled trials (RCTs) are the gold standard for the evaluation of healthcare interventions, but there is no one-size-fits all when it comes to the evaluation of medical technology. Especially in radiotherapy, where techniques, their users and implementation evolve over time, which makes that evidence from large, costly RCTs can quickly be outdated, even while the RCT is still ongoing.
Problem definition and aims
A patient’s toxicity risk can be predicted using normal tissue complication probability (NTCP) models. Via the model-based approach, NTCP models can be applied to select patients for a new treatment technique who potentially benefit most from it. In the Netherlands, the model-based approach is currently used to select patients for proton therapy across four cancer types. Despite the anticipated benefits of proton therapy in selected patients, the public and scientific (radiotherapy) community require evidence of the actual benefits of proton therapy in terms of toxicity reduction. To this extent, the model-based clinical evaluation has been introduced, in which observed toxicities in patients treated with a new therapy are compared to the predicted toxicity risks based on the standard therapy treatment plans (Figure 1). Despite being a seemingly efficient alternative to the RCT, the method is still novel and requires further methodological development. Our aim is to develop methods and provide guidance on the use of the model-based clinical evaluation for the evaluation of new radiotherapy techniques.
Plan of approach
We defined three objectives:
To investigate how missed confounders influence the validity of treatment effects estimated via model-based clinical evaluation.
To assess the impact of predictive performance (discrimination and calibration) of the NTCP models on estimates of the reduction of radiation-induced complications, and how to deal with this (e.g., model updating).
To compare how well an RCT and model-based clinical evaluation are able to provide valid treatment effect estimates under changing conditions over time (e.g., learning curve, different technique).
We will address these objectives in extensive simulation studies based on clinical data, simulating the whole process, from NTCP model development (considering confounders) to validation (assessing discrimination and calibration performance) to the model-based clinical evaluation (estimation of treatment effects). In these simulations we can manipulate different intermediate steps and measure the impact on the final observed treatment effect, allowing us to flexibly study the stated objectives and preconditions needed for valid treatment effect estimation using model-based clinical evaluation.
Envisaged results and impact
We will provide guidance on when and how the model-based approach can be used to evaluate treatment effects of new radiotherapy techniques. Considering nationwide discussions on the introduction of proton therapy in the Netherlands and the call for RCTs, our project addresses an important societal and scientific topic. Uptake of this guidance will be almost instantaneous due to our involvement in current prospective data registration programs, development of national indication protocols for proton therapy, and the evaluation of NTCP models and the clinical evaluation of the effects of proton therapy. Since all relevant stakeholders, including patients, clinicians, societies are involved in this project, methodology will be made fit for use, and fast dissemination and implementation of the results is expected.
Papers
2025
AMM Mulder, J Choi, LM Meijerink, W Van Amsterdam, AM Leeuwenberg, E Schuit, Considering causality in normal tissue complication probability model development: a literature review, July 2025 [medRxiv preprint]
Presentations
2025
Model-based clinical evaluation with counterfactual prediction vs. Enrichment trials: Assessing treatment effects as technologies evolve, MEMTAB 2025 [poster]
Missing confounding information in counterfactual prediction models: a simulation study on model-based treatment effect evaluation in radiotherapy, MEMTAB 2025 [poster]
Treatment effect estimation with counterfactual prediction using individual treatment plans: theory and application in radiotherapy, MEMTAB 2025 [poster]
Considering causality in normal tissue complication probability model development: a literature review, MEMTAB 2025 [poster]
Treatment effect estimation with counterfactual prediction using individual treatment plans: theory and application in radiotherapy, EUROCIM 2025 [poster]
Missing confounding information in counterfactual prediction models: an example of model-based clinical evaluation in comparison of radiotherapy techniques in cancer, The third “Causal Inference for AI” meeting, Erasmus MC, Rotterdam, Netherlands, February 10, 2025.
2024
The WhyMBA project, DUPROTON, 2024
Voorspelmodellen voor het personaliseren van behandelbeslissingen bijkankerpatiënten, UHD Lecture Ewoud Schuit (Youtube link)
2023
The WhyMBA project, Radiotherapy UMC Utrecht, 2023 [slides]