Translational barriers to digital twins in radiation oncology

Abstract

Digital twin research in radiation oncology has expanded rapidly across multiple domains, yet the field lacks definitional consensus and validated translational frameworks. A systematic search (PubMed, Scopus and Web of Science – September 2025) identified 903 records and six original studies met inclusion criteria. Appraisal of the available original studies revealed three recurring translational barriers: misuse of the term “digital twin” for virtual humans or patient-specific predictive models; overreliance on internal or in-silico validation; and limited benchmarking against clinically established alternatives. Progress toward clinical translation requires disciplined nomenclature, real-patient external validation, head-to-head benchmarking, explicit attention to data-pipeline and regulatory pathways.

Graphical abstract

Highlights

• •Systematic search yielded six radiation oncology digital twin studies included.
• •None of the included studies implements a true bidirectional closed-loop digital twin.
• •Three recurring barriers identified: terminology drift, validation gap, and benchmarking gap.

1. Introduction

The digital twin (DT) concept has attracted major attention in oncology as it promises a patient-specific computational representation that can anticipate treatment response and eventually support individualized decisions [1], [2], [3]. Radiation oncology has appeared especially well suited to that vision: it is data-rich, inherently quantitative, and already grounded in imaging, dosimetry, and radiobiological modeling [4], [5]. The National Cancer Institute has identified DTs in radiation oncology as a strategic funding priority [6], and early academic–industry collaborations are now underway [7]. Yet, enthusiasm alone does not constitute evidence. The label “digital twin” spans a remarkably wide range of methods, and that diversity can blur what is being claimed [8]. To address this issue, the National Academies of Sciences, Engineering, and Medicine (NASEM) recently called for clearer foundations in biomedical DT research [9], and recent oncology-specific perspectives emphasized the differences among DT and virtual human [10], [11].

In this study, we systematically reviewed the original radiation oncology DT literature with the aim of identifying the ontological differences and translational barriers across the available studies.

2. Material and methods

On 09 September 2025, PubMed, Scopus and Web of Science were searched using a structured query combining DT-related terms with radiation oncology terms. The full search string is provided in the Supplementary Material. PRISMA 2020 systematic review guidelines were followed for the present review [12]. After deduplication of records, 3 reviewers selected the manuscripts, performing an initial title and abstract review and a full text manuscript review. Original articles that explicitly reported the DT concept and applied to a radiotherapy context were included. Reviews, conference abstracts, and physics-only treatment-planning papers that did not use the term “digital twin” in title, abstract, or methods were excluded. The framework presented in Fig. 1 was developed by author consensus, mapping the foundational engineering definitions of DT [13], [14] together with the cardiology, immunology, and NASEM biomedical DT consensus reports [9], [15], [16] onto radiation oncology use cases.

Fig. 1.

Stages along the path to clinical digital twins in radiation oncology.

3. Results

The search returned 903 records (Scopus, n = 371; PubMed, n = 304; Web of Science, n = 228). After removal of 354 duplicates, 549 records were screened on title and abstract; 512 were excluded as irrelevant. Thirty-seven full-text articles were assessed for eligibility, of which 31 were excluded. Six original studies met all inclusion criteria [17], [18], [19], [20], [21], [22]. The PRISMA flow diagram is shown in Supplementary Fig. S1.

3.1. What counts as a digital twin?

In engineering and biomedical context, a DT is not any predictive model. The recent NASEM consensus on biomedical DT [9] and the oncology-focused perspective by Asghar and Chung [10] identify bidirectional integration as the main key distinguishing feature of a DT. In contrast, virtual humans are static or dynamically updated computational models of an individual patient with lack of a two-way integration in the clinical workflow. The core idea beneath DTs includes three fundamental components: specificity to an individual patient, the capacity to assimilate new data over time, and a closed feedback loop between the physical system and its digital counterpart [13], [14].

By that definition, radiation oncology literature is, at the present, mostly mislabeled. The included studies spanned patient-specific tumor-growth simulations calibrated on serial imaging [18], reinforcement-learning systems that optimize sequential treatment decisions [19], [20], a Bayesian optimization framework for dose-schedule personalization [21], a software platform demonstrated on synthetic data [22], and adaptive planning pipelines that select among precomputed plans using daily imaging [17]. These models are best described as virtual humans or patient-specific predictive models.

This is more than a semantic problem. If the same term is applied to describe a single-patient mechanistic simulation, a reinforcement-learning policy evaluated through a treatment simulator, a visual decision-support system, and a planning-stage automation tool, authors and reviewers are left without a shared evidentiary target, clinicians cannot judge what is being offered, and regulators cannot distinguish proof-of-concept from deployment-ready technology. Comparison across studies is also intrinsically difficult because these efforts are designed for different clinical questions, scales, and data types; the NASEM “fit-for-purpose” principle is helpful here [9]. Related disciplines have already encountered this challenge [15], [16].

A conservative way forward is to reserve the term “digital twin” for its strict, bidirectional definition and to describe earlier-stage work using more precise alternatives. Fig. 1 distinguishes three stages along the path to clinical DTs in radiation oncology: patient-specific predictive models calibrated retrospectively (Stage 1, virtual human); models with serial recalibration during care (Stage 2, adaptive virtual human); and true DTs embedded in the clinical workflow with bidirectional, automated data flow (Stage 3). Only Stage 3 meets the engineering definition; Stages 1 and 2 should not be marketed as DTs. On this framework, the current radiation oncology literature has remained almost entirely at Stage 1. A few studies have approximated Stage 2 behavior, by updating model parameters with intra-treatment imaging [18] or by selecting among precomputed plans using daily CBCT [17]. However, these demonstrations remained retrospective and no published radiation oncology study has reached Stage 3 yet. Importantly, this framework is taxonomic rather than hierarchical: a Stage 1 virtual human may be entirely appropriate for its intended clinical purpose and requires no closed-loop architecture to deliver value. The framework is intended to identify what model class is being described and what validation standard applies.

Stage 1 (virtual human, retrospective predictive model): patient-specific model calibrated on retrospective data and used offline. Stage 2 (adaptive virtual human): model with serial recalibration during care, but without automated, bidirectional integration into the clinical workflow. Stage 3 (digital twin): bidirectional, closed-loop system in which model outputs influence treatment decisions and subsequent patient data automatically update the model.

3.2. Virtual cohorts and simulators are useful, but they are not validation

One of the most valuable capabilities of DT methods is the generation of synthetic patient trajectories and the use of treatment simulators for counterfactual reasoning. Once a model is credible, virtual cohorts can support sensitivity analyses, regimen exploration, uncertainty quantification, and hypothesis generation for future trials [23]. The problem arises when those same virtual experiments are treated as evidence that the model is clinically accurate.

The current literature has illustrated the tension clearly. Some studies have used observed patient data to calibrate and test short-term predictions of tumor dynamics, an important early step toward real-world validation [18]. Optimized treatment pathways have been evaluated through treatment simulators rather than observed outcomes through those guidelines [19]. Dose-schedule optimization has been demonstrated on entirely virtual cohorts whose growth parameters were sampled from literature values [21], [22]. Even when real imaging data are used, evaluation has focused on plan-quality endpoints rather than clinical outcomes [17]. These studies are methodologically thoughtful, but model-generated or simulator-mediated outcomes establish internal consistency and scenario exploration, but do not establish external clinical validity.

Prediction-model methodology has distinguished this for decades: a model may be well calibrated to the data used to develop it (whether parameters are inferred from a population cohort or from a patient's own longitudinal data) and still fail to describe future dynamics on independent real patients [24], [25], [26]. When a model fitted on limited data is used to produce synthetic trajectories or simulate alternative regimens, the apparent precision of the simulation can easily exceed the evidentiary strength of the underlying data. Virtual cohort size is not a substitute for real-patient diversity. In silico trials and simulator-derived recommendations should therefore be presented as hypothesis-generating unless the underlying model has first demonstrated robust performance on independent real-world data.

A second, often neglected, point is that validating a true DT involves more than benchmarking the simulator. Because a Stage 3 DT, by definition, depends on automated data assimilation from the clinical workflow, validation must also evaluate the upstream pipeline: data acquisition cadence and reliability, sensor and image quality drift, latency between physical and virtual states, and integration with electronic health records, treatment planning systems, and image-guided radiotherapy platforms. A model whose simulator performs well in offline analysis can still fail clinically if the data infrastructure that feeds it does not.

3.3. Incremental value remains weakly benchmarked

A DT (or a candidate DT) must add something clinically useful that simpler approaches cannot provide. The added value may be trajectory forecasting, counterfactual treatment simulation, sequential decision support, or faster adaptive replanning. Regardless of the intended use, incremental value must be shown against a credible comparator.

Here, the literature has offered early but incomplete evidence. Personalized radiobiological models have been shown to outperform generic population models for longitudinal tumor volume prediction and early recurrence detection [18]. Adaptive planning frameworks have demonstrated improved organ of interest sparing relative to conventional clinical plans [17]. Yet across the broader DT corpus, formal head-to-head benchmarking against clinically established tools remains exceptional. Reinforcement-learning policies have been compared to observed clinical decisions, but not against validated clinical decision tools [19], [20]. Dose-schedule optimizations have been compared to standard-of-care dosing, but within simulation environments rather than against real clinical outcomes [21]. In no case has a DT been benchmarked against the best available alternative in a prospective clinical setting.

DTs will often be most attractive in settings where strong legacy models already exist. When a DT is intended to guide adaptation, de-intensification, or treatment sequencing, the relevant question is not whether it can generate plausible individualized curves, but whether it performs better, or yields better decisions, than the best available alternative. That comparison may involve discrimination, calibration, decision-curve analysis, or net clinical benefit rather than AUROC alone. Without a comparator, complexity can easily be mistaken for value [27].

4. Discussion

The barriers identified above are tractable, and most of the required solutions have already been articulated in the broader prediction-model and clinical-AI literature. We propose six minimum requirements for future radiation oncology DT studies, plus three further requirements that follow directly from the closed-loop nature of true DTs.

First, preregister model specifications (model architecture, training and validation data sources) in a public registry (e.g., OSF, ClinicalTrials.gov). Preregistration should specify comparator plans and validation endpoints, particularly when in silico trial claims are being made. Second, declare the intended use and maturity stage of the proposed model, distinguishing virtual-human or predictive-model precursors from candidate DTs. Third, report real-patient and virtual-patient sample sizes and population characteristics separately. Fourth, distinguish proof-of-concept demonstrations, internal validation, and external validation, and avoid presenting in silico cohort performance as evidence of clinical accuracy. Fifth, benchmark against at least one appropriate comparator (an NTCP model, a radiomic signature, a nomogram, or a transparent clinical baseline), matched to the claimed use case. Sixth, use existing reporting and appraisal tools such as TRIPOD+AI, PROBAST, and, when decision-support interfaces are being tested, DECIDE-AI or CLAIM [28], [29], [30], [31]. Dedicated DT-specific reporting standards will likely become necessary once the field reaches genuine Stage 3 deployments.

Beyond these prediction-model fundamentals, three further requirements are specific to true DTs. First, validate the data-assimilation pipeline as a system: input cadence and reliability, sensor and image-quality drift, latency between physical and virtual states, and behavior under pipeline degradation. Second, report computational latency relative to the clinical decision window the twin is intended to inform (daily replanning vs. on-table). Third, define a model-drift governance plan describing how the deployed twin will be monitored, when it will be retrained, and who is accountable for clinical performance over time.

Regulatory considerations cannot be deferred. A clinically deployed DT will fall under Software as a Medical Device frameworks: in the United States under the FDA SaMD risk categorization and recent guidance on AI/ML-enabled devices [32], and in Europe under the Medical Device Regulation in conjunction with the AI Act [33]. Because true DTs continuously assimilate new patient data and may update their behavior over time, they sit naturally within the FDA predetermined change control plan paradigm and require explicit specification of which model components are locked, which can adapt, and within what bounds. ISO/IEC 62304 software-lifecycle requirements and IMDRF SaMD risk principles also apply [34]. Authors of DT studies intended for clinical translation should describe, even briefly, how their proposed system would map onto these frameworks.

For models intended to influence treatment selection or adaptation, these are conditions under which clinicians, reviewers, and regulators can judge whether a twin is ready to inform patient care.

The promise of DT in radiation oncology remains genuine. Few specialties combine serial imaging, quantitative dose information, and computational culture as naturally as radiation oncology. Progress will come from disciplined nomenclature, real-patient external validation, credible benchmarking, and explicit attention to data infrastructure and regulatory pathways.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work, the authors used ChatGPT to improve readability and perform minor language editing.

CRediT authorship contribution statement

Federico Mastroleo: Writing – original draft, Methodology, Investigation, Formal analysis, Conceptualization. Mariana Borras-Osorio: Writing – review & editing, Resources, Data curation. Shiv P. Patel: Formal analysis, Data curation. David M. Routman: Writing – review & editing, Supervision. Doug J. Moseley: Writing – review & editing, Supervision. Satomi Shiraishi: Writing – review & editing, Supervision. Andrew Y.K. Foong: Writing – review & editing, Supervision. Mark R. Waddle: Writing – review & editing, Validation, Supervision, Resources, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.