Application of Operational Research Methodologies in Radiation Oncology Decision-Making: A Systematic Review

2.4. Data Extraction and Analysis

A pre-coded worksheet was then coded: author, year, analytical purpose, deployed OR framework, radiation-oncology use, and most important empirical outcome. The methodological integrity was assessed with the help of the Joanna Briggs Institute checklist that is specifically designed to be used in observational studies.

Two independent reviewers used a standardised extraction form to extract data. The variables that were extracted were the author, publication year, type of cancer, OR methodology, study objective, and key findings. The discrepancies were addressed by engaging in a discussion and arriving at a consensus. The quality of each contribution was then classified as high, moderate, or low-quality using criteria weighted in terms of explicitness, replicability, and specificity of the documented assumptions. As a result, narrative synthesis was conducted. The research was split into thematic areas.

Variations in sample size, tumour type, and patient level were dealt with by narrative stratification, summarising results by cancer site, dataset size, and methodological soundness of the study to ensure a fair evaluation.

The synthesis of recurrent patterns was made possible by this thematic structuring that concentrated on a methodical cross-domain synthesis. Moreover, the thematic areas included operational research involvement, comparative research methodology, and additional recognised constraints of the field of radiation oncology.

A total of 104 records were filtered out at the retrieval level based on the inaccessibility of full text, duplication in databases, or non-conformity to publication-type criteria, and 13 reports were filtered out post-full-text based on their lack of OR/AI methodology, lack of radiotherapy applications, or lack of extractable outcomes.

Articles were filtered out based on full text, were not in English, were multi-databases, and did not have radiation oncology as their focus, and articles that lacked an operational research approach.

2.5. Quality Assessment

The methodological quality was assessed independently by two reviewers using the Joanna Briggs Institute (JBI) critical appraisal checklist for observational studies and the ISPOR-SMDM modelling good research practices framework for modelling studies [8–10]. Research was divided into high, moderate, and low-quality studies on the grounds of transparency of methods, reproducibility, and quality of reporting [11–14]. Where there was an area of contention, consensus conference deliberation was used to find a solution (see the Appendices).

2.6. Ethical Consideration

As this study was based exclusively on previously published literature and did not involve human participants, animals, or identifiable patient data, formal ethical approval was not required

3. RESULTS

The review had seven articles published between 2015 and 2025 exploring the use of operational research (OR) and artificial intelligence (AI) in radiation oncology. Three papers were about optimisation and simulation methods to enhance radiotherapy workflow efficiency, and four papers were about AI-assisted clinical decision-support systems to improve adaptive radiotherapy and dose optimisation [11–13]. Together, these trials found an increase in scheduling efficiency, workflow management, treatment planning, and dose personalisation, but most studies were restricted by small sample sizes and single-centre designs [8–10].

3.1. Operational Efficiency and Resource Optimisation

This experiment measured the effect of the transformation of a mathematical operational research model in twin Dutch radiotherapy centers [11]. The centres achieved statistically significant schedule improvements post-implementation: average root-mean-square deviation of session initiation decreased by 51% (103.0 minutes to 50.4 minutes) in one site, and the difference in the number of gaps had been reduced by 72% (18 to 5) in the other site. In a similar line, the inefficiency of switching between linear accelerators, specifically identified as a workflow bottleneck, reduced to 71 and 43, respectively, in the original and alternative arrangement. Computations of productivity improved the distance between automated and manual modalities: The automated heuristic took 1-90 min to perform compared to the manual iteration, which was still ongoing and took 1.5 days.

Another work also focuses on computational efficiency: their experiments on five small-scale problems by systematic pruning of redundant constraints and targeted reformulation of objective functions also reported solution-time savings that were deemed to be substantial, and faster at generating feasible schedules [12]. Nevertheless, there are still no specific temporal measurements. The model was subsequently tested in a large-scale clinical setting where the model was found to be capable of assigning wards, equipment modalities and session frequencies in strict adherence to the operational and regulatory clinical constraints without going out of bounds.

One previous study has combined system dynamics constructs, together with Discrete-Event Simulation (DES), to study Computed Tomography (CT) operations to extract performance indicators that are sensitive to stochastic variation [13]. Simulations showed that inpatient sentinel examinations caused, on average, a 23% higher fraction of the full-use technologist occupancy than like outpatient studies, but outpatient non-contrast protocol used 63% less resource-minutes than the inpatient counterpart. Out-of-sample DES microsimulations of a dedicated outpatient service projected intensified operational loads on both radiography and assistant personnel, accompanied by reduced variability in ranges and lower median queue-holding periods than the in-sample DES. Thus, the quantitative balance lends empirical support to the deliberate separation of inpatient and outpatient referral paths. These findings identified workflow bottlenecks and demonstrated opportunities for operational improvement through the service redesign.

In concurrent redesign Randomised Controlled Trials (RCTs), the reductions correlated with a quasi-experimental protocol comprising iterative, cyclical recalibration of the executive predictive model, structured cognitive interaction with repercussive stakeholders, and provision of micro-redistributive adjustments to personnel allocator grids collectively, this intervention was associated with and accounted for the 51% reduction in the variance of completion-to-request lead times and the near-total eradication of linac-switching cascades [11]. Thus, a reduction in delays pre-emptively resolves temporal and spatial interactions across the iterative treatment-arrival network prior to any full-scale prototyping phase.

Complementarily, this study situated Soft Systems Methodology within a structured framework of stakeholder co-design that precedes discrete-event simulation trials of the emergent designs [13]. In the meantime, Emsamrit and Boonmee provide case studies where constraint-pruning heuristics are used on artificially small testing sets, thus preempting the need to engage in pragmatic validation milestones that are required before larger-scale clinical adoption. These studies indicate that the success of implementation is not just determined by the model performance but also stakeholder participation and integration into workflow.

3.2. AI-Driven Personalisation and Clinical Decision Support

In this paper, the paradigm of operational scheduling is applied to the area of personalised adaptive radiotherapy, where a comprehensive set of numeric measures are offered to evaluate the effectiveness of the model (8). Two potential trial datasets, initially containing 117 and 292 cases, were narrowed down to give final cohorts of 67 NSCLC patients (extracted neuroanatomical, dosimetric, and longitudinal features; 297 columns) and 71 HCC patients (extracted dosimetric contours, laboratory values, and outcome features; 110 columns), giving dense, labelled clinical inputs. The authors then created 10,000 virtual patients by synthesising clinical embeddings with a Generative Adversarial Network (GAN) to build a transfer-compatible augmented space for policy learning. The hybrid double graph neural network improved the prediction accuracy, reducing the RMSD from 0.97 to 0.61 Gy/fraction for NSCLC and from 4.75 to 2.96 Gy/fraction for HCC, respectively. Augmentative decision fidelity metrics indicate that the Adaptive Radiotherapy Clinical Decision Support Framework (ARCliDS) that uses double GNNs would classify 36% of NSCLC and 50% of HCC clinician-prescribed tactics as good when reproduced, and at the same time improve 74% of NSCLC and 30% of HCC decisions that had historically been considered poor. Taken together, these results show that the model had a significant level of agreement with the clinician-prescribed treatment decisions and enhanced the ratio of hitherto suboptimal decisions. Nevertheless, predictive performance was still not as high in Hepatocellular Carcinoma (HCC) compared to NSCLC, which means that further model refinement and validation are required [15].

Niraula et al. (2021) emphasised the nuanced hereditary and physiological influences that adjust the tolerance to ionising radiation and argued that the empirical dose schedules need adjustment throughout the radiotherapy process [8]. Their Quantum Deep Reinforcement Learning (qDRL) architecture allows the full leverage of quantum indeterminism by listing the uncertainty of each clinical choice in a figurative way and incorporating an extensive set of biological, physical and genomic data inputs. Basing its results on 67 patients with Non-Small Cell Lung Carcinoma (NSCLC) in the training set and testing the algorithm on an independent data set of 174 affected subjects, the ensemble reported a 10% improvement in clinically significant decision indices compared to those provided by trained clinicians over an equivalent observational reference time. The same group of 67 patients with NSCLC was then interrogated with a reinforcement learning compartment (DRL) to obtain a dose surface pre-empting and safely organising dose intensification [9]. The system showed the capability of predicting dose distribution and prescribing dose elevation strategies, the augmented intensity of which is safely within the modern constraints. The third quantitative elaboration by the study depicted the effectiveness of personalisation in the context of head-and-neck cancer, where the planning of the treatment by Volumetric Modulated Arc Radiotherapy (VMAT) is characterised by multifactorial geometrical and dose-volume constraints [14]. Their clinical decision support tool, which was trained using 276 institutional past plans, forecasted Organ-At-Risk (OAR) radiation dose; a combination of these forecasts with attending physician precepts resulted in hybrid dosimetry that reduced radiation exposure of non-target tissues and maintained target coverage goals.

A comparative study of blind and AI-assisted contouring assessed this interaction [8]. The results showed that the unassisted cohort was conservative, with an average of 0.1 Gy above the retrospective reference doses, but the AI-assisted group had a median of 0.59 Gy above the upper reference dose, indicating that the algorithm was pushing the cohort towards the clinically justifiable upper limit. Interestingly enough, this research study revealed the same median confidence rating of 6.9 on a 10-point scale of self-generated and AI-assisted recommendations, which suggests that doctors viewed the model as equally trustworthy and reliable. Sher et al. (2021) approached human–machine interplay by permitting clinicians to issue a draft recommendation, followed by an AI supplying an evidence-based dose estimate. The Henderson–Doctors Directive (HDD) took the conservative option, resulting in greater sparing of Organs At Risk (OAR) by 4.3–16 Gy compared with the standalone clinician recommendation and by 5.6–9.1 Gy beyond the AI output in isolation. This complementary architecture was extended in the study as the variability of uncertainty in the clinical context was modelled with the help of parameterised quantum states, allowing clinician-like variability to exist in the loop of reinforcement learning [9]. This approach enabled the model to better account for the uncertainty inherent in clinical decision-making.

The paper reported that the dose adjustments due to AI were about 10% higher than when they were changed by clinical instructions without the use of AI, where fidelity was assessed using Root Mean Square Error (RMSE) calibrated to historical judgement [8]. In a previous study [9], Click or tap here to enter text. dose variability was further narrowed, resulting in an AI prescription of 0.59 Gy (SD 0.78) greater than the unguided clinical prescription, and AI-generated treatment contours were off-target by 0.3 Gy (SD 1.2) from retrospective standards. The size of the physician-based deviation and the comparatively smaller deviation margin left to the AI provide circumstantial evidence of augmented operators to favour safe, graded dose escalation, which increases the therapeutic ratio. The highest dosimetric evidence was provided in a study, which stated that deep-learning plans attained 22 to 75% of Organ-At-Risk (OAR) dose limits even with an intervention of over 3 Gy, which pushed clinical plans to a net average of 4.3 to 16 Gy reduction in OAR dosimetric [14]. The decrement thresholds are in line with the clinical consensus, as decrements exceed 3 Gy have been defined to be in the category of producing an observable clinical response. Taken as a bloc, these dosimetric observations provide sound, clinically meaningful empirical evidence that AI-inefficient workflows create empirically observable and clinically useful improvement, and in this notice the areas that exist, which, as opposed to theoretical benefit, come down to measurable outcomes protocol, which are not only statistically and physiologically measurable.

Niraula et al. (2021) introduced qDRL, which considers the uncertainty of decisions by encoding them in quantum states, and exhibited training on both a Qiskit simulator and an IBM quantum computer [8]. This method has the theoretical and practical ability to apply large, high-dimensional metastatic data-scapes using quantum reinforcement learning. On the other hand, Niraula et al. [9] based their work on the elements that facilitate clinicians to adopt the technology and created ARCliDS based on microservices composed of a Python server and a reactive R Shiny visualisation. The resulting platform wraps any complexity of programming in a request-response paradigm and therefore makes the modelling predictions accessible to any oncologist with entry-level scripting experience. A more traditional supervised paradigm was used by Sher et al. (2021), who operationalised and prospectively assessed the framework within an individually approved care pathway by the IRB, obtained consent for their study, and continuously tracked intervention effects. The model ensured epidemiologic validity and physician oversight by incorporating GBT into live tumour boards [14]. Together, the qDRL complex learning handling, user-centred pipeline, which offers a coherent feedback channel, and the workflow integrated with the IRB make sure that learning systems move beyond curiosity and towards patient-safe, demonstrably relevant, scalable, and clinically relevant oncology environments.

3.3. Implementation Challenges and Future Translation

A pervasive quantitative limitation arises from the narrowly defined sample scope. The study enhancements were confirmed within single-week planning horizons across only two centres [11]. Another study validation examined five small test cases and real-world examples [12]. The study employed a discrete-event simulation restricted to the computed tomography load of a single hospital [13]. The study’s reinforcement learning agents obtained a training set of 10,000 synthetic patients, and the original clinical records were evidently insufficient for convergence [12]. Hence, although the observed percentage gains and reductions in the root-mean-square deviation appear promising, the external validity is curtailed by the combined constraints of minimal multi-centre observation and synthetic data augmentation. Improvements, both in magnitude and statistical significance, remain convincing within their respective environments, yet mandate multi-site, prospective clinical trials to verify their extension beyond the observed settings.

The evidence highlights that there is a 51% reduction in schedule standard reduction and almost complete removal of linac switch events; second, the brittleness of the model is assessed with discrete-event simulation, which works with workload redistribution, where inpatient-related staffing pressure increases by 23 percentage points; and third, patient-level AI-assisted dosing algorithms are trained, where the reliably decreased root-mean-square deviation in fractionated dose (0.97 to 0.61 Gy) satisfies clinical significance criteria.

Building trust and creating massive adoption remain some of the bottlenecks of health informatics. In an observational study, clinicians reported that the reliability of the AI-generated recommendations was rated as equal to the personal expert judgment system, but the systems made practitioners quietly persuade them to make the choices that best correspond to evidence-based best practice [10]. The importance of trust was supported in this study, as the providers regularly incorporated AI-generated prognostic probabilities into shared care pathways[14]. In order to embody that confidence, patterns of human-like decision uncertainty can be detected in a quantum inference architecture, an explicit attempt to demystify the margin of error of the machine. [9]. Several studies emphasised transparency and clinician trust as important factors influencing AI adoption in clinical practice.

Table 1 summarises the objectives, methods, interventions, and principal findings of the seven studies. Table 2 compares the operational or AI method, evidence base, validation approach, and clinical relevance of the included studies. The detailed numerical screening counts are presented in Table 3.

4. DISCUSSION

The convergence of Operational Research (OR), Artificial Intelligence (AI), and Machine Learning (ML) in the radiation oncology field has been gaining momentum in the last five years. The analysed articles all reflect a shift from traditional rule-based radiotherapy processes to AI-supported and data-driven decision-support systems. Applications include optimisation and discrete-event simulation models, reinforcement learning and quantum-inspired AI frameworks. In the literature, these methods were related to the refinements in treatment planning, workflow, resource allocation, and decision-making in adaptive radiotherapy.

4.1. Interpretation and Comparison with Previous Studies

These results align with the results of Krishnamurthy et al. [27], who placed AI as a tool for decreasing manual labor and enhancing the utilisation of resources in the processes of radiotherapy. They also build on McCullum et al. [5], who discovered in their systematic review that most Markov models compare fixed treatment policies, and there is limited application of optimal sequential decision models. The current review introduces the evidence based on appointment scheduling, service redesign, adaptive dosing and treatment planning. Thus, OR makes contributions both at the system level and at the patient level, not differentiating between operational efficiency and clinical adaptation as two different issues.

The pattern of the results must be interpreted more cautiously as compared to broad technology-centred reviews. Duke and Papanikolaou [20] mentioned AI use in imaging, planning, quality assurance, delivery, and adaptive radiotherapy, but Hoebers et al. [30] noted that models that have an immediate impact on a patient were less developed compared to efficiency applications. The imbalance was similar in the seven studies included: scheduling and workflow models provided quantifiable operational improvements, whereas some clinical decision-support models used limited, single-centre, or even artificial datasets. The paramount concern is thus no longer technical feasibility, but the reproducibility, safety, and utility of the performance across institutions.

A juxtaposition of the analyses conducted by [11, 12] depicts conclusive effects of Operations Research (OR) models on the efficiency of radiotherapy scheduling. The results of Vieira’s implementation showed quantifiable benefits that prove that optimisation frameworks can be implemented in non-simulation settings. Emsamrit and Boonmee enhanced the computational efficiency by re-writing the optimisation constraints and testing the model on a number of test cases. Their results validate the computational tractability principle: the run-time and practicable optimality of allocations decreases are gradual to the clinical processes and applied interoperability of scheduling models.

However, article [13] pursued the implementation of Operations Research (OR) in resource provisioning by combining both system dynamics and Discrete-Event Simulation (DES) models to assess the Computed Tomography (CT) waiting lists and the utilisation of shared resources. Their empirical results show that inpatient examinations consume fewer total staff hours than outpatient examinations. By modelling stochastic demand and intrinsic variability, OR is more test-fidel, with respect to traditional deterministic optimisation, allowing decision-makers to evaluate and optimise alternative service structures along planning horizons. Their findings suggest that separating inpatient and outpatient CT services may improve resource utilisation and reduce waiting times. Used alongside contemporary scheduling optimisation studies, the integration of system dynamics and DES within the capacity planning context exemplifies the coherent treatment of OR techniques at micro (near-real-time scheduling of examinations) and macro (long-term placement of resources and staff capacity) analytic horizons within consolidated oncology services.

These innovations can only be diffused effectively as four parallel activities, namely: continuous statistical cross-validation, cross-disciplinary co-design of workflow contours, gradual cross-site trials that convert numerical advantage into long-term clinical leverage, and governance structures that maintain adaptive thresholds to reassess.

The Niraula et al. (2021–2023) investigation represents a paradigm shift to adaptive and personalised treatment planning. Their 2021 publication unveiled a quantum Deep Reinforcement Learning (qDRL) architecture engineered to encode dose-response relations from a triad of genotypic, clinical, and historical dosimetric variables. The model was designed to account for uncertainty in clinical decision-making by incorporating probabilistic representations of the patient and treatment variables. There was an increase in the accuracy of treatment recommendations in the cross-validation of Non-Small Cell Lung Carcinoma (NSCLC) cases in a cohort. This article presents a plan for integrating patient-specific heterogeneity in radiotherapy dosimetry processes systematically. Within [9], the authors presented ARCliDS, a commercially viable architecture that integrates predictive supervised learning (ARTE) with policy-optimising Deep Reinforcement Learning (DRL) in a continuous clinical human interface. Performance measures were not limited to accuracy, as they included quantitative measures of clinician interaction and trust. The observed clinician acceptance of AI-assisted recommendations suggests that trust and interpretability may be as important as predictive accuracy for successful implementation of AI-assisted recommendations. It is interesting to note that the perceived trust in AI outputs (6.9/10 Likert scale) was as high as the perceived trust in unaided clinical recommendations, which means that clinicians will be able to operate with convenience soon when interpretable systems with operational transparency are involved.

Conversely, study [10, 12] built upon reinforcement learning methods, which incorporated Graph Neural Networks (GNNs) to adapt radiotherapy to patients with NSCLC and HCC, and its overall importance is that it has shown that AI-based systems can aid personalised adaptation of treatment. These results indicate that reinforcement-based learning and graph-based AI might enhance the stability of dose suggestions and enable adaptive decision-making in RT. Nevertheless, there are significant concerns about the generalisability and clinical performance in practice when relying on synthetic training data, which makes it necessary to verify it with prospective data on various clinical datasets in the future [10].

Nevertheless, in a past study [14], complementary use of artificial intelligence was reported with Volumetric-Modulated Arc-Therapy (VMAT) for head and neck malignancy. Their directive aid instrument with institutional treatment plan informed it, which was prospectively scrutinised on patients. The interconnection between human and neural directives in a hybrid manner reduced the OAR median dose by more than 3 Gy in the trajectories and by 4.3 to 16 Gy in redistribution compared to human directives only. This simultaneous improvement of the volumes of contours confirms the idea that AI systems, when integrated into clinical practice, progressively and quantitatively improve dosimetric outcomes, thereby providing viable and clinically significant benefits to the population of patients [16–18].

The analysed literature presented a number of strengths, such as methods innovations, quantifiable changes in clinical or operational results, and greater AI integration in clinical processes [8, 9]. Nevertheless, critical flaws still exist, such as small single-centre cohorts, the use of synthetic datasets, a lack of prospective validation, and a lack of evidence in the context of long-term clinical implementation [10, 14].

4.2. Practical Implications for Radiation Oncology

The results favor a gradual implementation. Scheduling and simulation systems could initially run in shadow mode with the current rosters and future tracking of waiting time, variability in start time, machine switching, overtime, cancellations and treatment interruption. Adaptive treatment tools must have extra controls, such as clinician override, explicit uncertainty display, prospective audit, subgroup performance checks, data drift monitoring, and recorded responsibility for the final decision. Radiation oncologists, medical physicists, radiographers, operations researchers, data scientists and service managers should be included as implementation teams. These protections convert the numerical performance into responsible workflow modifications.

On the service level, OR models can be used to inform the division of the inpatient and outpatient routes, workforce, and the planning of the linear accelerator capacity. Nevertheless, local optimisation can move the workload to another group of staff or prioritise urgent and complex cases. As such, implementation must consider efficiency, access, safety, staff burden, patient experience, and clinical outcomes. When data and technical support are limited, the simpler and more transparent models can be used in resource-constrained centres. Only when such complex AI systems show added value compared to traditional optimisation, and can be sustained within the local governance arrangements, should these AI systems be adopted.

4.3. Innovation and Contribution

This review uses standard OR and AI-enabled models around real radiation oncology decisions, in contrast to earlier reviews, which focused on one model family or general AI functions. It maps seven studies in the domains of scheduling, service configuration, adaptive dosing, and treatment planning; compares their validation and clinical integration; and finds a common translation gap between model performance and sustained use. This decision-centred synthesis clarifies where evidence may support controlled implementation and where multi-centre prospective validation remains necessary. Its main contribution is an integrated account of operational and clinical decision-making, rather than a catalogue of algorithms.

LIMITATIONS AND FUTURE TRANSLATION

Although the examined literature collectively affirms the value of operational Even though the reviewed literature, on the whole, supports the importance of Operational Research (OR) and Artificial Intelligence (AI) in the process of decision-making refinement in the field of radiation oncology, there are a number of relevant limitations to the generalisability of the studies in question and, by extension, its future to find direct implementation in clinical practise. A sample size is another weakness, as it is rather limited, and the scope of data is not diverse enough [19, 20]. In ref. [8], successful results were found using reinforcement learning and ARCliDS architectures in cohorts consisting of only 67 patients with Non-Small-Cell Lung Cancer (NSCLC), but ref. [14] applied their prospective decision-support intervention to a sample of 50 patients. Narrow disease centres, monolithic care environments, and small study centres all intrinsically truncate wider extrapolations to diffuse, heterogeneous populations of patients [21]. Attempts to counter these limitations have been unsuccessful. In ref. [8], limited external validation was conducted in a group of 174 patients, but the external dataset was limited to only NSCLC. This equivocal external examination also eliminates inferences towards the classification of a range of malignancies usually encountered in clinical radiation oncology [22]. One partial methodological weakness that has persisted is the use of synthetic datasets to offset clinical sample size shortages [23, 24]. In a study, 10,000 artificial records generated by a generative adversarial architecture were used to guide reinforcement learning algorithms [10]. In the absence of satisfactory numerical performance, the rationalisation of predictive generalisability is conditional on the similarity between artificial patterns and the complex, high-dimensional variances of real clinical groups. The limitations include articles only in English, a small number of included studies, heterogeneity of methodologies, and the inability to perform meta-analysis. The lack of longitudinal, multi-institutional datasets increases the risk of overfitting and limits generalisability to diverse clinical populations [25]. The practicality and computational efficiency of the model is one of the parallel market mismatches that are yet to be tackled. This study [11] demonstrated that the outpatient scheduling performance was significantly improved, but it took longer and more repetitive re-specification of local procedural tenets in four clinical settings, indicating a volatile transfer between different organisational infrastructures. As noted in a previous study [12], the findings were related to marginal reductions in central processing overhead, and formalisation was limited to five simulated circuits plus one retrospective example, limiting the external validity in oncology settings in which the number of patients was in the hundreds and treatment journeys were complex and connected.

Therefore, the published advances are still tentative standards as opposed to strong and extensively generalisable algorithms [26].

Another major challenge was trust and subsequent adoption by stakeholders. Even though [9] recorded the average physician self-assurance in AI of 6.9/10, which is similar to self-assessment, only one radiation oncologist evaluated eight cases. Systematic multicentre questionnaires in different fields are justified to understand whether AI prescriptions will be considered authoritative and implemented in everyday practice in full. It was shown in a prior study [14] that a hybrid AI-physician planning modality outperformed either of the components, but longitudinal buy-in and scaled workflow embedding have not been reported. The external validity of AI depends on limited and artificial data, which implies that the advantages achieved may not be the entire package concerning general and real-world groups of patients [27–29].

Therefore, the current systematic review does not address any of the paths to formal approval or the necessities of the explainability of algorithmic outputs. In total, the available investigations find that improvements in throughput, tailored interventions and clinical outcomes are attained. Its use is, however, limited by its small cohort size, use of artificially generated data, lack of external validation, use of clinician belief and inefficient regulatory systems [30].

There are a number of limitations to this review. The studies included only those that were published in English, which could have led to a language bias. There were only a few eligible studies, as the application of OR and AI in radiation oncology centers is still young. There was significant heterogeneity in study design, methods, outcome measures and validation techniques that precluded quantitative meta-analysis and restricted direct comparisons across the studies [14]. Moreover, it is not possible to rule out publication bias because studies with positive results have a greater chance of being published compared to those with negative or inconclusive results.