Proton Therapy's Hidden Uncertainty: New "Ensemble" Models Aim to Protect the Brainstem and Spinal Cord

Table of Contents

• Key Points
• Why Proton Therapy and RBE Matter
• Why the Brainstem and Spinal Cord Need Special Focus
• Study Design: Building an Ensemble Model From Published Research
• The Literature Search: Finding the Right Studies
• Quality Assessment and Data Extraction
• Treatment Planning Simulation: Testing the Models
• How the Ensemble Weights Were Calculated
• Key Findings: What the Models Showed
• Clinical Implications: What This Means for Patients
• Limitations: What This Study Could Not Prove
• Recommendations and Future Directions
• Frequently Asked Questions
• Source Information

Key Points

• Proton therapy uses a fixed RBE of 1.1, but real RBE varies with energy and tissue type.
• Brainstem and spinal cord are serial organs; small dose errors can cause severe complications.
• New ensemble models eRBE-B and eRBE-SC combine three RBE models weighted by study quality and patient numbers.
• For the brainstem, the ensemble reduced dose underestimation by about 1–2% compared with the Carabe model alone.
• This proof-of-concept study needs validation with real patient data before clinical use.

Why Proton Therapy and RBE Matter

Proton therapy is a form of radiation treatment that has become a cornerstone of modern radiation oncology. Unlike traditional X-rays, protons deposit most of their energy at a specific depth in the body, known as the Bragg peak, allowing doctors to target tumors with high precision while sparing nearby healthy tissues.

A critical concept in proton therapy is the relative biological effectiveness (RBE), a measurement of how powerful proton radiation is at damaging cancer cells or healthy tissue compared with conventional photon radiation (X-rays). For decades, most clinics have used a fixed RBE value of 1.1, a simplification recommended by early research such as that of Paganetti and colleagues. This means doctors assumed protons are 10% more damaging than X-rays, regardless of the situation.

That simplification has a catch. RBE is not actually constant. It depends on linear energy transfer (LET), which describes how much energy the beam deposits along its path, and on the tissue-specific radiosensitivity (measured by the alpha/beta ratio, or α/β), which describes how sensitive a particular tissue is to radiation dose. Higher LET values generally increase biological damage, while tissues with a low α/β ratio are especially responsive to those effects.

Several research groups have developed variable RBE models to capture this complexity, including the Carabe, Wedenberg, and McNamara models. These are so-called phenomenological models based on the linear-quadratic framework, and they account for LET and α/β values. Yet the models often disagree with each other because of differences in data fitting, underlying assumptions, and regression methods. Studies by Rørvik et al. and McMahon et al. have highlighted these inconsistencies. With no direct RBE measurements or robust clinical benchmarks available, clinicians face significant uncertainty when choosing which model to trust.

Why the Brainstem and Spinal Cord Need Special Focus

The brainstem and spinal cord are known in radiobiology as serial organs. Like a chain where one weak link breaks the whole structure, damage to even a small focal segment of these organs can cause significant clinical problems.

For patients with head and neck cancers, these organs at risk (OARs) often sit very close to the clinical target volume (CTV), the area that needs the radiation. Even small errors in calculating the RBE-weighted dose can lead to major differences in treatment outcomes:

• Excessive radiation to the brainstem can cause severe or even fatal complications.
• Overexposure of the spinal cord may result in motor dysfunction (problems with movement) or sensory deficits (numbness or loss of feeling).

This is why accurate dose estimation for these structures is essential. The biologically effective dose is especially uncertain near the distal edge of the Bragg peak, where LET values change sharply. According to QUANTEC guidelines, the brainstem's maximum tolerated dose (D_max) should remain within 54–59 Gy (EQD2, the equivalent dose in 2-Gy fractions used to standardize treatment schemes). Exposures of 64–65 Gy (RBE) or higher carry a high risk of necrosis (tissue death) or cranial neuropathy (nerve damage).

Study Design: Building an Ensemble Model From Published Research

Researchers face a fundamental obstacle when studying RBE: there are no direct patient-level RBE measurements, and no harmonized quantitative endpoints across studies. That rules out a traditional meta-analysis, which statistically combines comparable numerical outcomes.

Instead, the team turned to meta-synthesis (MS), an approach that aggregates patterns of model usage, methodological rigor, and theoretical foundations across heterogeneous published studies. Unlike meta-analysis, MS provides a structured but flexible framework for combining diverse types of evidence. Using this method, the researchers developed two new ensemble models:

• eRBE-B for the brainstem
• eRBE-SC for the spinal cord

The study was a proof-of-concept project. Its goal was to demonstrate that combining the RBE-model literature into an ensemble is feasible and useful, not to directly validate RBE values against clinical benchmarks. As the authors state plainly, the eRBE models are probabilistic tools designed to address uncertainty, not replacements for clinical validation.

The Literature Search: Finding the Right Studies

The research team conducted a systematic search of three major databases: Web of Science, PubMed, and Scopus. The search covered studies published up to July 2023 and used a four-tier screening strategy:

1. First tier: Studies had to focus on cancer patients and specifically address brainstem and spinal cord dose assessments.
2. Second tier: Only studies from established peer-reviewed collections were included.
3. Third and fourth tiers: Studies had to incorporate both a fixed RBE of 1.1 and variable RBE models in their research design.

For the brainstem, for example, the search combined terms such as "Brainstem" OR "Medulla oblongata" OR "Pons" OR "Midbrain" with "proton therapy" and a series of RBE-related keywords. The PRISMA guidelines (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) were used to document each screening step, including title, abstract, and full-text review. The protocol was registered in the INPLASY database under registration number INPLASY202540057.

Quality Assessment and Data Extraction

Each included study was scored with the Joanna Briggs Institute (JBI) critical appraisal checklist for quasi-experimental studies. The checklist has 9 questions; every "yes" answer earned 1 point, for a maximum score of 9.

Because the data sources were published papers rather than human assessors, traditional inter-assessor calibration did not apply. Instead, the researchers handled potential variability with a structured weighting system built on two criteria:

• Methodological rigor, assessed by the JBI score
• Empirical strength, reflected by the number of cases in each study

These two metrics were combined and normalized to produce ensemble weights for each RBE model. Three pieces of data were extracted from each study: the specific RBE models used, the JBI quality score, and the number of cases evaluated.

Treatment Planning Simulation: Testing the Models

To compare how the different RBE models perform in a realistic treatment scenario, the researchers created treatment plans using the RayStation 8B treatment planning system (RaySearch Lab, Stockholm, Sweden). These plans were then imported into MCsquare (Université catholique de Louvain, Belgium), a Monte Carlo simulation toolkit written in Python that simulates how radiation particles interact with tissue.

A commercial Rando phantom (The Phantom Laboratory, Salem, NY, USA) was used to model the anatomy of a head and neck cancer patient. The Rando phantom is widely used in radiation oncology research because its tissues closely mimic real human anatomy.

To minimize radiation exposure to critical structures, two proton beams were applied at gantry angles of 90° and 270°, replicating a typical clinical treatment strategy. The prescribed dose to the CTV was 60 Gy (RBE), which matches real-world dosing for head and neck tumors.

MCsquare produced spatial distribution files for dose and LET. From these, the team calculated RBE-weighted doses using both the published single models and the new ensemble models. The final biologically weighted dose files were analyzed in 3D Slicer software, which generated dose-volume histograms (DVHs) for the CTV, brainstem, spinal cord, and parotid gland. A dose-volume histogram is a graph that shows how much radiation each organ receives.

The team also ran calculations for the parotid gland (α/β = 3), which showed moderate sensitivity with minimal impact from variable RBE modeling in this study.

How the Ensemble Weights Were Calculated

The ensemble models combine the Carabe, Wedenberg, and McNamara models. These three were chosen because of their well-documented foundations in the linear-quadratic framework and their extensive use in previous research. Each model incorporates dose-averaged LET (LET_d) to account for how proton beams interact with tissue. The authors note that newer databases such as the Particle Irradiation Data Ensemble (PIDE) exist, but they deliberately focused on these three models for their established reliability.

The weighting formula uses two terms for each model: the study quality scores (s_f) and the case numbers (n_f), each divided by the cumulative totals across all studies (S and N). In plain language, a model earns more weight when it is backed by high-quality studies and large numbers of patients.

Here is the concrete example given in the paper, using five hypothetical studies:

• Quality scores for the five studies: 8, 9, 7, 8, and 9 (total S = 41).
• Patient counts: 4, 5, 6, 7, and 8 (total N = 30).

The Carabe model was used in 3 studies with quality scores 8, 7, and 9, and case numbers 4, 6, and 8 (s_f = 24, n_f = 18). This produced a raw weight of about 0.59. The Wedenberg model appeared in 1 study with a quality score of 8 and 4 cases, giving a raw weight of about 0.16. The McNamara model was used in all 5 studies, with the full quality score of 41 and all 30 cases, giving a raw weight of 1.0.

The total weight was 0.59 + 0.16 + 1.0 = 1.75. After normalization so the weights sum to 1:

• Carabe: 0.59 / 1.75 ≈ 0.34
• Wedenberg: 0.16 / 1.75 ≈ 0.09
• McNamara: 1.0 / 1.75 ≈ 0.57

The resulting ensemble model formula is: eRBE = 0.34 × Carabe + 0.09 × Wedenberg + 0.57 × McNamara.

The weighting process also accounts for biological factors. The eRBE value decreases as the proton fraction dose (D_p) increases, meaning biological effects are less pronounced at higher doses. A low α/β ratio amplifies the effects of LET_d and D_p, adding tissue-specific sensitivity. In the McNamara model, LET_d modifies biological effects through a square-root dependency, which adds mathematical complexity. The essence of eRBE is simple: combine three empirical formulas through weighted averaging, balancing the individual biases of each model while capturing their strengths.

Key Findings: What the Models Showed

The headline result concerns the brainstem. Compared with the Carabe model alone, the eRBE-B model reduced dose underestimation, with a mean dose difference of −0.89 Gy (−1.44%) and a maximum dose difference of −1.08 Gy (−1.61%).

Interestingly, eRBE-B closely aligned with the McNamara model for the brainstem, yielding a mean dose (D_mean) of 9.88 Gy and a maximum dose (D_max) of 64.07 Gy. This alignment makes sense given that McNamara received the largest ensemble weight (0.57).

Across the critical organs studied, the ensemble models showed enhanced consistency and stability compared with any single model. They mitigated the biases inherent to single-model approaches, particularly in the high-LET regions near the distal edge of the Bragg peak.

Another important finding concerns the standard fixed RBE of 1.1. The analysis, including Monte Carlo simulations and DVH comparisons, revealed that fixed RBE models may underestimate biologically effective doses by up to +10% in these sensitive regions. When a brainstem is already close to its safety limit, that 10% could be the difference between a safe plan and a harmful one.

Toward the end of the spine analysis, the parotid gland (α/β = 3) demonstrated only moderate sensitivity, with minimal impact observed from variable RBE modeling. This suggests that some organs are less affected by RBE choices than the brainstem and spinal cord.

Clinical Implications: What This Means for Patients

For patients with tumors near the brainstem or spinal cord, this research matters because safety margins are tight. The QUANTEC guidelines state that the brainstem's maximum tolerated dose should not exceed 54–59 Gy (EQD2), while exposures of 64–65 Gy (RBE) or higher pose a high risk of necrosis or cranial neuropathy.

The eRBE approach offers several potential advantages for clinical care:

• Better biological dose accuracy: The ensemble model reduces the underestimation seen with fixed RBE (1.1) models, especially in high-LET regions such as the brainstem.
• Organ-specific adjustments: By incorporating tissue-specific radiosensitivity (α/β) and LET distributions, eRBE models can be tailored to individual critical structures.
• Earlier identification of risk: By capturing deviations that fixed models miss, eRBE models may help identify cases where safety thresholds could be exceeded, improving patient-specific risk assessment.
• More consistent planning: Because the ensemble balances the biases of three established models, it is less vulnerable to the weaknesses of any single theoretical approach.

The authors emphasize that the ensemble framework is biologically adaptive rather than assuming a constant RBE. This is a step toward uncertainty-aware treatment planning, where doctors know not just the dose but also how much confidence to place in that number.

Limitations: What This Study Could Not Prove

The authors are careful to describe this as an exploratory proof-of-concept study. Several limitations deserve attention:

• No direct clinical validation: The eRBE models were not validated against real patient outcomes or direct RBE measurements. They consolidate existing knowledge rather than proving which model is biologically correct.
• No direct experimental phantom validation: Although the MCsquare Monte Carlo toolkit has been extensively benchmarked against clinical measurements in previous studies, the current work did not perform its own phantom experiments.
• Meta-synthesis, not meta-analysis: The MS approach aggregates patterns of model usage and study characteristics rather than statistically integrating validated experimental results. It is therefore more exploratory than confirmatory.
• Limited model selection: Only three models (Carabe, Wedenberg, McNamara) were blended. Newer datasets such as the PIDE database were not incorporated.
• Simulation-based testing: The dose calculations used a Rando phantom, not actual patient images or treatment histories.

Recommendations and Future Directions

For patients and clinicians, the takeaway is not that eRBE models are ready to replace standard practice. Instead, they represent an interim solution to current uncertainties while more rigorous tools are developed.

The authors identify clear next steps:

1. Test the eRBE approach with real patient data to confirm its safety and effectiveness.
2. Validate the models using alternative simulation platforms such as TOPAS and FLUKA, which are other Monte Carlo toolkits.
3. Establish standardized RBE validation protocols through future clinical and experimental efforts.
4. Continue developing uncertainty-aware RBE modeling that quantifies confidence alongside dose estimates.

For now, the eRBE methodology shows promise as a next-generation computational tool for improving the precision and biological relevance of proton therapy treatment planning. Patients receiving proton therapy for head and neck cancers can expect that researchers are actively working to make dose calculations more accurate and safer — especially for the delicate structures at the base of the brain and along the spinal cord.

Frequently Asked Questions

What is proton therapy and why is its biological effect uncertain?

Proton therapy targets tumors precisely using protons that stop at a specific depth. Its biological strength, called RBE, is often assumed to be 1.1, but RBE actually varies with energy deposited and tissue sensitivity. This uncertainty matters most for sensitive organs like the brainstem and spinal cord, where even small dose errors can be harmful.

Why are the brainstem and spinal cord especially at risk during proton therapy?

The brainstem and spinal cord are serial organs, meaning damage to a small segment can cause serious problems. For head and neck cancers, these structures sit close to the tumor. Guidelines say the brainstem's maximum tolerated dose is 54–59 Gy, while 64–65 Gy carries high risk of tissue death or nerve damage, so accurate dosing is critical.

What are the new 'ensemble' RBE models eRBE-B and eRBE-SC?

These are models that combine predictions from three established RBE models: Carabe, Wedenberg, and McNamara. Each is weighted by study quality and patient numbers. eRBE-B is for the brainstem and eRBE-SC for the spinal cord. They aim to reduce dose underestimation and give radiation oncologists a more safety-conscious way to plan treatment.

How much does the ensemble model reduce dose underestimation compared with a single model?

For the brainstem, the eRBE-B model reduced dose underestimation compared with the Carabe model alone. The mean dose difference was about -0.89 Gy, or -1.44%, and the maximum difference was about -1.08 Gy, or -1.61%. These are average differences from modeling, not guarantees for individual patients.

Is the fixed RBE of 1.1 still accurate for proton therapy?

The article says fixed RBE of 1.1 may underestimate biologically effective doses by up to 10% in sensitive regions near the brainstem and spinal cord. For a brainstem already close to its safety limit, that 10% could mean the difference between a safe plan and a harmful one. More accurate models are being developed.

Will these ensemble models replace standard proton therapy planning soon?

No. The study is a proof-of-concept, not ready for clinical use. The models were not validated against real patient outcomes or direct measurements. They are probabilistic tools to address uncertainty. Next steps include testing with real patient data and validating using other simulation platforms before they can be considered for routine practice.

What are the limitations of this study on RBE models?

There was no direct clinical validation and no experimental phantom validation in this work. It used meta-synthesis, not meta-analysis, and only blended three models. Testing used a phantom, not actual patient images. The authors describe it as exploratory, and newer datasets like PIDE were not included.

Should I get a second opinion on my proton therapy plan if my tumor is near the brainstem or spinal cord?

Yes, if your tumor is close to the brainstem or spinal cord, a second opinion focused on the proton therapy plan can be valuable. Standard planning uses a fixed RBE of 1.1, which may underestimate biologically effective dose by up to 10% near the end of the proton beam. This matters because brainstem tolerance is around 54–59 Gy (EQD2), and doses above 64–65 Gy carry high risk of serious damage. Newer ensemble models are not yet clinically validated, but a second opinion can review whether your plan accounts for RBE uncertainty. Diagnostic Detectives Network provides independent expert second opinions.

Source Information

Original article title: Ensemble RBE Modeling in Proton Therapy: A Meta-Synthesis Framework for Dose Assessment in the Brainstem and Spinal Cord.

Authors: Lee SH, Chao PJ, Wu ZJ, Wu JJ, Shiau J, Shih HH, Lee TF.

Affiliations: National Kaohsiung University of Science and Technology; Linkou Chang Gung Memorial Hospital and Chang Gung University College of Medicine; Kaohsiung Chang Gung Memorial Hospital and Chang Gung University College of Medicine; Kaohsiung Medical University, Taiwan.

Publication details: Cancer Control, Volume 32, pages 1–18, 2025. Received February 12, 2025; revised June 10, 2025; accepted June 15, 2025. DOI: 10.1177/10732748251355396. Open access under a Creative Commons Attribution-NonCommercial 4.0 License.

Note: This patient-friendly article is based on peer-reviewed research. It is intended for educational purposes and is not a substitute for professional medical advice. Patients should discuss their radiation treatment plans with their oncology care team.