From Semantic Modeling to Precision Radiotherapy: An AI Framework Linking Radiobiology, Oncology, and Public Health Integration

1. Introduction

In the past decades, the number of scientific research studies has grown exponentially, creating vast reservoirs of knowledge in many fields [1,2,3,4]. However, a significant fraction of that knowledge remains underutilized because it is scattered in databases and challenging to access or interpret, especially in very specialized areas [5,6,7,8,9].

This study is based on a straightforward premise: the combination of literature mining and artificial intelligence (AI) can revolutionize the mapping of the thematic evolution and conceptual foundations of radiotherapy [10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28], radiobiology [29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52] and oncology [53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72], because by enabling automated, scalable and interpretable analyses, this approach overcomes the limitations of traditional searches and manual curation, providing a broader, more current and more interdisciplinary vision of the state of the art [7,73,74,75].

Traditional bibliometric tools are often unable to capture the subtle patterns that matter in specialized areas, thus highlighting the value of AI-based approaches that exploit metadata, semantic models and co-occurrence networks; therefore, the use of unsupervised machine learning and text mining in our previous work revealed emerging trends and knowledge gaps in areas such as nanocomposites, controlled drug delivery, photovoltaics, catalysis, biosynthesis and spectroscopy, showing that computational approaches can accelerate the synthesis and application of scientific knowledge [76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93].

In this work, we apply this strategy to an interdisciplinary corpus of thousands of articles indexed in Scopus, PubMed and Web of Science, intending to rigorously structure, validate and interpreting how radiotherapy, radiobiology and oncology have consolidated as an interconnected domain from 1964 to 2025, because to our knowledge, this is the first comprehensive effort to apply AI-based semantic and temporal analysis to the literature of these disciplines at this scale and depth.

None of these studies has considered both the semantic and temporal dimensions of the literature by a unified analysis pipeline, and in particular, there is a gap in the methodological literature between bibliometric/quantitative approaches and clinically oriented predictive models. The rest of the literature is organized around four main strands, each one focused on a particular aspect of the problem, including bibliometric and visualization studies, which have been conducted by using tools such as VOSviewer and CiteSpace for analyzing the evolution of keywords and co-authorship networks. For instance, Wang et al. (2024) [94] reports a keyword co-occurrence network, which shows the main trends in the field from 2014 to 2025, however, these studies are essentially descriptive, and in general, they do not include topic modeling, deep semantic analysis, or time series methods.

Methodological proposals, such as the one by Andrei and Arandjelovic (2016) [95], use hierarchical Dirichlet processes and temporal similarity graphs to model the evolution of topics over time, but this framework has not been used in the biomedical context nor in radiotherapy. Meanwhile, AI for radiotherapy literature, including studies by Tabibi et al. (2025) [96] and Lastrucci et al. (2024) [97], use deep learning to predict clinical outcomes and refine treatment strategies in radiotherapy, consequently, precision and efficacy of the treatment have increased significantly, although these studies predict clinical endpoints, whereas the mining or semantic representation of the scientific literature itself has not been addressed.

Reviews, such as Momin et al. (2020) [98] and Trifiletti and Showalter (2015) [99], discuss the integration of clinical and genomic data in radiotherapy, but do not perform a semantic and temporal analysis of the scientific output, therefore, to summarize, these literature strands suggest a methodological gap between purely bibliometric studies and clinically oriented predictive models. To the best of our knowledge, there is no published study that performs an integrated analysis of combining semantic modeling (e.g., topic modeling, co-occurrence networks) with structured temporal analysis (e.g., historical time series, change point detection) in a unified framework for the radiotherapy, radiobiology, and oncology literature, which justifies the novelty of the present study, that aims at synthesizing bibliometric indicators, semantic modeling, and temporal dynamics in a single framework providing both analytical insight and predictive perspective on the scientific evolution of the field.

By integrating major databases and processing more than three thousand unique documents with advanced computational modeling, the study goes beyond traditional bibliometrics, providing a dynamic and interpretable view of how these fields have evolved over six decades, and linking molecular mechanisms, clinical practice, and technological development within a single analytical framework. In so doing, the study demonstrates a new approach to scientific progress: artificial intelligence as a tool for synthesis and interpretation, rather than allowing valuable findings to remain scattered and fragmented. AI can facilitate their transformation into cohesive insights that can inform translational advances and support more precise, efficient, and equitable healthcare.

2. Methods

The search strategies applied to Scopus, PubMed, and Web of Science were carefully harmonized to identify publications addressing the intersection of radiotherapy, radiobiology, and oncology, with the guiding question: “What are the key connections and advances at the interface of radiotherapy, radiobiology, and oncology as reflected in the recent scientific literature?” The query used in Scopus was TITLE-ABS-KEY ((radiotherapy) AND (radiobiology OR "radiation biology") AND (oncology)); in PubMed, a combination of MeSH terms and free-text fields was applied—("Radiotherapy"[MeSH Terms] OR "Radiotherapy"[All Fields]) AND ("Radiobiology"[All Fields] OR "Radiation Biology"[All Fields]) AND ("Oncology"[MeSH Terms] OR "Oncology"[All Fields]); and in Web of Science, the direct key (Radiotherapy AND Radiobiology AND Oncology) was used. These strategies initially retrieved 2,507 records from PubMed, 741 from Scopus, and 714 from Web of Science, yielding a total of 3,962 documents.

3. Results

The structured computational pipeline began with the automated ingestion of files from Scopus, PubMed, and Web of Science. We standardized and normalized the core data fields, and we retained additional metadata for each data source. We merged the data, and when the same publication was included twice, we selected the record with the most complete data, therefore we removed 594 duplicate publications based on ID numbers and 25 based on titles. This resulted in a total of 3,343 publications, and thus the final dataset was composed of a large number of unique publications. The majority of the publications were retrieved from PubMed (66.23%), followed by Scopus (21.96%) and Web of Science (11.82%), and so the distribution of publications across these databases was not uniform.

Figure 1 provides a graphical synthesis of the results, and in the upper part of the figure, the number of publications per year is presented, which shows three phases: low publication rate from 1964 to 1990, moderate growth from 1991 to 2010, and rapid growth from 2011 onward, because this coincides with the rise of stereotactic body radiotherapy, protons, and prediction models. The years 2020 and 2023 had the highest number of publications (~240), but the decrease in 2025 is probably due to incomplete indexing of the literature, and thus this decrease may not reflect the actual number of publications. The word cloud at the bottom of the figure was generated based on titles and abstracts, and common terms such as radiat, cell, treatment, cancer, dose, patient, tumor, radiotherapi, fraction, therap, and radiobiolog, are highlighted.

Additional terms, such as proton, irradi, brachytherapy, repair, dna, toxicity, and response, refer to molecular mechanisms and emerging techniques, while surviv, risk, model, and outcome emphasize prognosis, and clinic, trial, meta, and evalu indicate the prevalence of translational trials.

The lower panel depicts the top ten most frequent terms across years, and cell is the most frequent term, with >550 occurrences in 2020, which reflects the high number of cellular/molecular investigations, and terms such as radiat, patient, dose, cancer, treatment, tumor, and radiotherapi increased in parallel between 2015 and 2023, because they are closely related to the main topics of the field. Effect and clinic increased gradually, reflecting increasing efforts on therapeutic effects and clinics, and thus they are important for the development of the field.

Collectively, this indicates limited vocabulary use until the early 1990s, followed by gradual increase, and a sharp increase after 2010, which correlates with the development and consolidation of the field, and so the field has undergone significant changes over the years. Following text preprocessing, Latent Dirichlet Allocation (LDA) was performed using k = 10 topics, with lemmatization, biomedical stopword filtering, and frequency trimming (no_below=2–5, no_above=0.5) to ensure vocabulary stability. Models were trained separately for titles and abstracts (passes = 10, random_state = 42), and validated through inspection of word distributions and stability across runs. The top 20 LDA-derived keywords were subsequently used to construct the co-occurrence supergraph, providing a quantitative foundation for thematic mapping. Titles included head and neck, breast, and prostate cancer; adverse effects; image-guided planning; and highly conformal treatment, while abstracts included DNA damage and repair; dose-response modeling; clinical trials; immunoradiotherapy; and normal tissue adverse effects, therefore the topics covered a wide range of subjects. For each topic cluster, word association maps were created using the 20 most probable words, and these maps were integrated into a single large map, which thus provided a comprehensive overview of the field. Main nodes were radiotherapy, tumor, DNA repair, and adverse effects, and all files (CSV, SVG, PNG) were generated for visualization in Gephi and topological analysis, therefore the results can be easily visualized and analyzed. Yearly word counts and smoothed frequency curves indicated sustained growth in the most prominent biology and technology concepts, which is a significant finding.

This has led to an increasingly interdisciplinary field linking radiotherapy, radiobiology and oncology, with reciprocal interactions between technological developments, cellular processes and clinical practice, and so the field has become more complex and multidisciplinary. The approach emphasizes repeatability, transparency and reproducibility of methods to ensure the interpretability and verifiability of findings, therefore it is essential for the development of the field. Briefly, LDA was applied to the titles and abstracts of the 3,343 non-duplicate publications, identifying 10 topics for each corpus, reflecting the broad scope of a literature that encompasses radiotherapy, radiobiology and oncology, and full counts and model parameters are presented in Equations 1-20 of the Supplementary Information, which thus provides a detailed description of the methodology and results.

In the titles, topics ranged from clinical and anatomical foci to molecular mechanisms and technical approaches. Topic_0t, dominated by cancer, radiotherapy, patient, head, neck, breast, and prostate, underscores the concentration on specific tumor types treated with radiotherapy, including advanced cases and toxicity. Topic_1t highlights radiobiology, clinical, and oncology, indicating the integration of biological foundations with clinical practice. Topic_2t clusters cancer, breast, carcinoma, prostate, and esophageal, characterizing comparative studies. Topic_3t emphasizes radiation, oncology, biology, and molecular, reflecting interest in the mechanisms of radiation action. Topic_4t, centered on stereotactic, body, radiosurgery, and lung, captures the SBRT literature in pulmonary neoplasms. Topic_5t addresses radiation, beam, ion, and proton, underscoring dose delivery physics and tissue protection. Topic_6t, with tumor, brain, model, and imaging, points to modeling and preclinical studies of brain tumors. Topic_7t organizes tumor, cell, DNA, repair, and pathway, reflecting the molecular biology of DNA damage. Topic_8t, dominated by cell, human, expression, and gene, refers to in vitro experimentation. Finally, Topic_9t, with dose, brachytherapy, model, and radiobiological, is associated with dose modeling and brachytherapy.

In the abstracts, methodological granularity is greater. Topic_0a, with model, dose, imaging, and flash, suggests imaging-based modeling and planning for FLASH-RT. Topic_1a, centered on cell, tumor, DNA, repair, and damage, focuses on the biological mechanisms of radiation. Topic_2a, with proton, ion, RBE, and particle, delineates the literature on particle therapies. Topic_3a, clustering patient, survival, RT, and surgery, reflects prognostic clinical studies. Topic_4a, with toxicity, breast, risk, Gy, and SBRT, addresses toxicity in breast cancer patients treated with precision radiotherapy. Topic_5a, highlighting trial, immunotherapy, preclinical, and targeted, points to the rise of immunoradiotherapy and combination therapies. Topic_6a, with expression, gene, protein, and blood, characterizes research on molecular biomarkers. Topic_7a, with dose, Gy, plan, and mouse, refers to preclinical trials and validation in animal models. Topic_8a, dominated by generic terms such as clinical, therapy, oncology, and development, likely reflects institutional or editorial texts. Finally, Topic_9a, with dose, fraction, tissue, and effect, emphasizes studies on dose fractionation in normal tissues.

The thematic landscape emerging from our analyses is two-dimensional. It consists of interdependent clinical-anatomical and mechanistic-molecular axes, because the clinical-anatomical axis encompasses disease sites, therapeutic modalities, and patient-centered outcomes, including survival and toxicity. In contrast, the mechanistic-molecular axis encompasses cellular responses to ionizing radiation, DNA damage and repair pathways, gene expression programs, and biomarker development. The overlap of the axes is not coincidental; rather, it represents a truly translational ecosystem where conceptual and experimental advances are translated into clinical decision-making, and, reciprocally, clinical needs drive mechanistic inquiry. Consequently, several novel fronts have been gathering steam within this matrix, including ultrafast radiotherapy (FLASH-RT), immunoradiotherapy, individualized predictive modeling, and the routine inclusion of molecular biomarkers. Collectively, these fronts constitute a precise, forward-looking map of how the field is maturing and where it is headed; therefore, the co-occurrence supergraph (Fig. 1 inset) denotes the semantic backbone of the literature encompassing radiotherapy, radiobiology, and oncology.

In the graph, nodes represent keywords extracted from titles, abstracts, and descriptors, while edges encode the frequency of co-occurrence. Edges are visually weighted according to connection strength: thicker, reddish edges correspond to stronger semantic relations (as among cancer, radiotherapy, tumor, and particle), while thinner, bluish edges correspond to weaker associations. From a topological perspective, the network is dense and cohesive, displaying a density of 0.5, a diameter of 1, an average degree of 9.5, and an average weighted degree of 1962.2, reflecting intense co-occurrence relationships. A mean clustering coefficient of 0.5 with 1140 triangles reveals substantial local cohesion, while a low modularity of 0.02 and a single weakly connected component indicate a highly integrated thematic space.

All network metrics were computed in Gephi using a directed, weighted setup, which means the numeric ranges are shifted, and hence why diameter = 1 and density = 0.5 correspond to the directed adjacency definition and not a fully connected graph, because the weighted degree is a count of total co-occurrence, which is why these values are higher than unweighted degrees. Together with low modularity and high clustering, the metrics paint a picture of a tight but non-trivial lexical network, i.e. evidence of thematic convergence without the network collapsing into a single blob. However, when we re-ran the analysis on an undirected, thresholded version of the data, we found the relative ordering of important nodes and communities to be essentially unchanged, with a Pearson correlation coefficient greater than 0.9, which therefore supports the robustness of our interpretation.

From a practical point of view, the literature has converged into a single, integrated domain rather than a collection of disconnected subfields; thus, it is well positioned for efficient knowledge transfer. The primary node in this topology is cancer, closely connected to tumor, radiotherapy, particle, and breast, while technical subfields (proton, stereotactic, imaging) are peripherally located yet act as anchors for technological and methodological development.

When mapped onto term frequency curves over time, the topology confirms an increasing frequency of terms starting in the 2000s and accelerating after 2010, mirroring the overall surge in biomedical research. Key terms (cell, cancer, dose, treatment, radiotherapy) have expanded sharply, confirming their central role. The consistency between temporal dynamics and network topology reinforces the centrality of cancer and radiotherapy, while highlighting the growing prominence of molecular response terms (DNA, repair), clinical outcome terms (survival, toxicity), and technical parameters (dose, fraction).

In aggregate, this literature is characterized by strong interdependencies among clinical, molecular, and technological concepts, without rigid semantic boundaries. Therefore, this consolidation implies efficient knowledge transfer across scales—from DNA repair pathways to dose planning and patient outcomes. Practically, the supergraph provides a decision-ready analytical tool to frame systematic reviews, identify underexplored clusters, and guide targeted literature mining strategies in oncological radiotherapy. Thus, the semantic cohesion of the field is not only descriptive but functional, furnishing a robust and verifiable backbone for a continuously learning, precision-oriented radiotherapy ecosystem capable of integrating emerging innovations such as FLASH-RT, immunoradiotherapy, and biomarker-driven personalization.

From this integration, four main hypotheses emerge: (1) cancer and radiotherapy function as structuring axes of recent scientific production; (2) the growth of terms such as cell, dose, effect, and treatment reflects the emphasis on therapeutic personalization, mechanisms of action, and clinical efficacy; (3) the co-occurrence of tumor, DNA, repair, and survival indicates intensified research in precision medicine and response biomarkers; and (4) the densely connected network confirms the transversal nature of the field, integrating molecular biology, medical physics, and clinical practice within a unified scientific ecology.

We iterated these hypotheses to inform the search strategies used to locate research at the interface of radiotherapy, radiobiology, and oncology. In Scopus, we searched titles, abstracts, and author keywords for records that are limited to biomedical subject areas, and from post-2014, which retrieved 25 records, and in PubMed, we combined MeSH with broad field searches and limited results to 2015-2024, and to clinical or translational publication types, which retrieved 10 records. Together, the refined search strategies maximised specificity and translational relevance to precision radiotherapy, and thus limited the resulting corpus to studies that are most likely to inform precision radiotherapy. In Web of Science, the query applied to the abstract field yielded 35 documents. After automated deduplication by DOI and title, the records were consolidated into a corpus of 61 articles, representing an analytical set centered on the molecular, therapeutic, and clinical trends of contemporary radiobiological oncology. This corpus serves to empirically validate the thematic and structural hypotheses identified through semantic analyses, providing a solid foundation for the investigation of high-impact case studies.

Semantic analysis of the corpus by Latent Dirichlet Allocation (LDA) reveals a three-part thematic structure, and when combined, topics extracted from titles collapse into three separate themes. The first theme is translational, with terms such as radiotherapy, effect, cell, combination, metastatic, and clinical, and corresponds to work that bridges combined therapeutic approaches in cellular systems with clinical trials. The second theme is biomarkers, with terms such as cancer, patient, trial, biomarkers, hypoxia, and radiosensitivity, and corresponds to work that bridges molecular signatures and microenvironmental factors with outcomes such as response and toxicity, while the third theme is mechanistic, with terms such as model, tissue, flash, normal, genomics, and learning, and corresponds to work that applies radiobiological modeling and machine learning to predict tissue responses, including under ultrafast dose-rate conditions, such as FLASH radiotherapy. Together, these three themes suggest a field that seamlessly bridges laboratory findings, molecular profiling, and clinical practice, while also developing data-driven approaches to predict and improve patient outcomes.

The analysis of abstracts confirms this tripartite thematic segmentation while adding a more refined methodological layer. Terms such as response, damage, DNA, repair, radiosensitivity, and cellular reinforce the focus on molecular responses to radiation, whereas words like parameter, model, radiobiological, high, and biological indicate the development and calibration of quantitative models. At the same time, expressions such as dose, risk, volume, and fractionation point to the optimization of dosing schemes. The consistent presence of signature, hypoxia, biomarker, and median cohort reveals the adoption of genomic signatures and clinical cohort analyses as predictive tools.

Co-occurrence analysis of the top 20 seed terms with the highest probability, such as prediction, parameter, radiobiology, vitro, biology, personalized, protocol, effect, response, and cancer, revealed densely connected graphs, and terms like response, cancer, normal, and radiobiology were found to be high-centrality bridge terms that linked clinical themes, including patient outcomes, protocols, and therapeutic effects, and mechanistic themes, including cellular responses, radiobiological parameters, and in vitro evidence. In a practical sense, these nodes are semantic hinges that orient experimental modeling and biological insight towards clinical application, thereby weaving together disparate strands of the literature into a cohesive translational narrative.

To refine the selection of the most relevant articles for case studies, priority was given to expressions that simultaneously encompass molecular dimensions, clinical applications, and therapeutic innovation. Terms in titles and abstracts corresponding to the translational and mechanistic cores of the field (e.g. biomarker signature, radiosensitivity, DNA repair, hypoxia-induced, FLASH radiotherapy, dose-response model, genomic classifier, precision oncology, radioresistance, therapeutic window, combined modality, clinical trial phase, translational framework, machine-learning prediction, brachytherapy dose escalation) were combined, so as to reflect the temporal evolution of the frequency curves and the density of the semantic supergraph, thus ensuring that the extracted literature remains closely connected to the thematic axes generated by the computational analysis, and this strategy prevented the search from spreading away from the literature that bridges molecular mechanisms and clinical decision-making as well as technological implementation.

A total of 28 out of the 61 documents met the predefined criteria, namely those that mentioned in their titles or abstracts at least one of the fifteen recommended key expressions: biomarker signature, radiosensitivity, DNA repair, hypoxia-induced, flash radiotherapy, dose–response model, genomic classifier, precision oncology, radioresistance, therapeutic window, combined modality, clinical trial phase, translational framework, machine-learning prediction, or brachytherapy dose escalation. Documents lacking a DOI or equivalent identifier, as well as those outside the 2014–2025 timeframe, were excluded from the analysis. Table 1 presents the selected documents, indicating year of publication, identified terms, and the corresponding numerical classes assigned according to the established thematic axes [37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62].

Screening was conducted using a Python script that loaded the consolidated records, removed duplicates, and systematically examined the title and abstract fields. The algorithm applied a lexical search function, converting texts to lowercase and checking, term by term, for the exact presence of expressions defined from LDA modeling and the most relevant co-occurrence patterns. For each record, lists of identified terms were generated and subsequently unified into a new column, enabling structured thematic analysis. The documents were then classified into one or more of the five main classes: DNA repair and molecular response, precision oncology and genomic models, individual radiosensitivity, tumor radioresistance, and emerging technologies in radiotherapy.

Each document group was analyzed individually using the PaperProcessor [100] script, which performs semantic extraction and summarization guided by large language models (LLMs). The process is directed by the subject parameter, which steers the thematic interpretation of each textual segment. To ensure analytical coherence and conceptual focus, the subject was adapted to the nature of each document class. In the first four classes, focused on molecular mechanisms, genomic stratification, clinical variability, and tumor resistance, open and interpretative prompts were used to generate explanatory syntheses, mechanistic articulations, and conceptual inferences. The fifth class, dedicated to technological innovation in areas such as FLASH radiotherapy, heavy charged particles, and voxel-level analytics, was conceived to clarify the underlying physical principles, the design of delivery systems, and the first steps toward clinical application. This structure enabled a balanced and interpretative reading of the literature supported by artificial intelligence.

Even though large language models (LLMs) are central to the semantic synthesis stage, their use within the PaperProcessor pipeline is carefully bounded by independent preprocessing, normalization, and statistical validation steps. This layered design ensures methodological transparency: all prompts are fixed and stored in the source code, allowing for external auditing and exact reproducibility of each run. Each model inference is time-stamped and logged to a CSV file to create a complete auditable record of all outputs, and at the same time, an independent unsupervised topic modeling check using Latent Dirichlet Allocation (LDA) is performed to provide confirmation that the themes are consistent across all the documents. Together, these steps help minimize interpretive bias and ensure that the AI synthesis is evidence-based, transparent, and reproducible.

Table 1. Selected studies based on key expressions and thematic classification.

#	DOI	Title	Reference	Year	Terms Found	Classes *
1	10.1007/s12553-024-00895-y	Current trends and future perspectives in hadron therapy: radiobiology	[101]	2024	DNA repair	[1]
2	10.3389/fonc.2021.687672	The Time for Chronotherapy in Radiation Oncology	[102]	2021	DNA repair	[1]
3	10.1667/RADE-24-00037.1	What's changed in 75 years of RadRes? - An australian perspective on selected topics	[103]	2024	DNA repair	[1]
4	10.3389/fnano.2025.1603334	Radiobiological perspective on metrics for quantifying dose enhancement effects of High-Z nanoparticles	[104]	2025	DNA repair, precision oncology	[1, 2]
5	10.1016/j.clon.2021.09.003	Can Rational Combination of Ultra-high Dose Rate FLASH Radiotherapy with Immunotherapy Provide a Novel Approach to Cancer Treatment?	[105]	2021	flash radiotherapy	[5]
6	10.1016/j.canlet.2025.217895	Unraveling the dual nature of FLASH radiotherapy: From normal tissue sparing to tumor control	[15]	2025	flash radiotherapy, radiosensitivity	[5, 3]
7	10.18632/oncotarget.10996	Next generation multi-scale biophysical characterization of high precision cancer particle radiotherapy using clinical proton, helium-, carbon- and oxygen ion beams	[106]	2016	precision oncology	[2]
8	10.3390/ijms26136375	Radiation Therapy Personalization in Cancer Treatment: Strategies and Perspectives	[107]	2025	precision oncology	[2]
9	10.1080/14737140.2018.1458615	The effects of radiotherapy on the survival of patients with unresectable non-small cell lung cancer	[108]	2018	precision oncology	[2]
10	10.1016/j.hoc.2019.07.001	Toward a New Framework for Clinical Radiation Biology	[109]	2019	precision oncology	[2]
11	10.1016/S0007-4551(18)30383-7	Kidney cancer and radiotherapy: Radioresistance and beyond	[110]	2018	radioresistance	[4]
12	10.1093/jrr/rrz048	Carbon-ion irradiation overcomes HPV-integration/E2 gene-disruption induced radioresistance of cervical keratinocytes	[111]	2019	radioresistance, radiosensitivity	[4, 3]
13	10.1016/j.radonc.2022.08.001	Accurate prediction of long-term risk of biochemical failure after salvage radiotherapy including the impact of pelvic node irradiation	[112]	2022	radiosensitivity	[3]
14	10.1186/s13014-024-02566-8	Alpha/beta values in pediatric medulloblastoma: implications for tailored approaches in radiation oncology	[113]	2025	radiosensitivity	[3]
15	10.32471/exp-oncology.2312-8852.vol-43-no-3.16554	Can SARS-CoV-2 change individual radiation sensitivity of the patients recovered from COVID-19? (experimental and theoretical background).	[114]	2021	radiosensitivity	[3]
16	10.3390/cells12030360	Comparison of Radiation Response between 2D and 3D Cell Culture Models of Different Human Cancer Cell Lines	[115]	2023	radiosensitivity	[3]
17	10.2298/SARH220131085N	Current aspects of radiobiology in modern radiotherapy – our clinical experience	[116]	2022	radiosensitivity	[3]
18	10.1002/mp.13751	Genomics models in radiotherapy: From mechanistic to machine learning	[117]	2020	radiosensitivity	[3]
19	10.1016/j.rpor.2014.11.006	Hypersensitivity to chemoradiation in FANCA carrier with cervical carcinoma-A case report and review of the literature.	[118]	2015	radiosensitivity	[3]
20	10.3390/cancers8030028	Paradigm Shift in Radiation Biology/Radiation Oncology Exploitation of the H2O2 Effect for Radiotherapy Using Low-LET (Linear Energy Transfer) Radiation such as X-rays and High-Energy Electrons	[119]	2016	radiosensitivity	[3]
21	10.1259/bjr.20170949	Radiation biology and oncology in the genomic era	[120]	2018	radiosensitivity	[3]
22	10.1088/1361-6560/aac814	Radiobiological parameters in a tumour control probability model for prostate cancer LDR brachytherapy	[121]	2018	radiosensitivity	[3]
23	10.1136/bmjopen-2021-059345	The COMPLETE trial: HolistiC early respOnse assessMent for oroPharyngeaL cancEr paTiEnts; Protocol for an observational study	[122]	2022	radiosensitivity	[3]
24	10.1016/j.radonc.2023.109868	Voxel-based analysis: Roadmap for clinical translation	[123]	2023	radiosensitivity	[3]
25	10.1002/cnr2.1126	Progressive breast fibrosis caused by extreme radiosensitivity: Oncocytogenetic diagnosis and treatment by reconstructive flap surgery.	[124]	2019	radiosensitivity, DNA repair	[3, 1]
26	10.1002/mp.15177	Radiobiological comparison between Cobalt-60 and Iridium-192 high-dose-rate brachytherapy sources: Part I—cervical cancer	[125]	2021	radiosensitivity, hypoxia-induced	[3, 5]
27	10.1016/j.ijrobp.2015.11.013	Influence of Nucleoshuttling of the ATM Protein in the Healthy Tissues Response to Radiation Therapy: Toward a Molecular Classification of Human Radiosensitivity	[126]	2016	radiosensitivity, radioresistance	[3, 4]
28	10.3390/proteomes13020025	Deciphering Radiotherapy Resistance: A Proteomic Perspective	[127]	2025	radiosensitivity, radioresistance, DNA repair, precision oncology	[3, 4, 1, 2]

^{* Classes: (1) DNA repair, (2) Precision oncology, (3) Radiosensitivity, (4) Radioresistance, and (5) Flash radiotherapy.}

Table 1. Selected studies based on key expressions and thematic classification.

#	DOI	Title	Reference	Year	Terms Found	Classes *
1	10.1007/s12553-024-00895-y	Current trends and future perspectives in hadron therapy: radiobiology	[101]	2024	DNA repair	[1]
2	10.3389/fonc.2021.687672	The Time for Chronotherapy in Radiation Oncology	[102]	2021	DNA repair	[1]
3	10.1667/RADE-24-00037.1	What's changed in 75 years of RadRes? - An australian perspective on selected topics	[103]	2024	DNA repair	[1]
4	10.3389/fnano.2025.1603334	Radiobiological perspective on metrics for quantifying dose enhancement effects of High-Z nanoparticles	[104]	2025	DNA repair, precision oncology	[1, 2]
5	10.1016/j.clon.2021.09.003	Can Rational Combination of Ultra-high Dose Rate FLASH Radiotherapy with Immunotherapy Provide a Novel Approach to Cancer Treatment?	[105]	2021	flash radiotherapy	[5]
6	10.1016/j.canlet.2025.217895	Unraveling the dual nature of FLASH radiotherapy: From normal tissue sparing to tumor control	[15]	2025	flash radiotherapy, radiosensitivity	[5, 3]
7	10.18632/oncotarget.10996	Next generation multi-scale biophysical characterization of high precision cancer particle radiotherapy using clinical proton, helium-, carbon- and oxygen ion beams	[106]	2016	precision oncology	[2]
8	10.3390/ijms26136375	Radiation Therapy Personalization in Cancer Treatment: Strategies and Perspectives	[107]	2025	precision oncology	[2]
9	10.1080/14737140.2018.1458615	The effects of radiotherapy on the survival of patients with unresectable non-small cell lung cancer	[108]	2018	precision oncology	[2]
10	10.1016/j.hoc.2019.07.001	Toward a New Framework for Clinical Radiation Biology	[109]	2019	precision oncology	[2]
11	10.1016/S0007-4551(18)30383-7	Kidney cancer and radiotherapy: Radioresistance and beyond	[110]	2018	radioresistance	[4]
12	10.1093/jrr/rrz048	Carbon-ion irradiation overcomes HPV-integration/E2 gene-disruption induced radioresistance of cervical keratinocytes	[111]	2019	radioresistance, radiosensitivity	[4, 3]
13	10.1016/j.radonc.2022.08.001	Accurate prediction of long-term risk of biochemical failure after salvage radiotherapy including the impact of pelvic node irradiation	[112]	2022	radiosensitivity	[3]
14	10.1186/s13014-024-02566-8	Alpha/beta values in pediatric medulloblastoma: implications for tailored approaches in radiation oncology	[113]	2025	radiosensitivity	[3]
15	10.32471/exp-oncology.2312-8852.vol-43-no-3.16554	Can SARS-CoV-2 change individual radiation sensitivity of the patients recovered from COVID-19? (experimental and theoretical background).	[114]	2021	radiosensitivity	[3]
16	10.3390/cells12030360	Comparison of Radiation Response between 2D and 3D Cell Culture Models of Different Human Cancer Cell Lines	[115]	2023	radiosensitivity	[3]
17	10.2298/SARH220131085N	Current aspects of radiobiology in modern radiotherapy – our clinical experience	[116]	2022	radiosensitivity	[3]
18	10.1002/mp.13751	Genomics models in radiotherapy: From mechanistic to machine learning	[117]	2020	radiosensitivity	[3]
19	10.1016/j.rpor.2014.11.006	Hypersensitivity to chemoradiation in FANCA carrier with cervical carcinoma-A case report and review of the literature.	[118]	2015	radiosensitivity	[3]
20	10.3390/cancers8030028	Paradigm Shift in Radiation Biology/Radiation Oncology Exploitation of the H2O2 Effect for Radiotherapy Using Low-LET (Linear Energy Transfer) Radiation such as X-rays and High-Energy Electrons	[119]	2016	radiosensitivity	[3]
21	10.1259/bjr.20170949	Radiation biology and oncology in the genomic era	[120]	2018	radiosensitivity	[3]
22	10.1088/1361-6560/aac814	Radiobiological parameters in a tumour control probability model for prostate cancer LDR brachytherapy	[121]	2018	radiosensitivity	[3]
23	10.1136/bmjopen-2021-059345	The COMPLETE trial: HolistiC early respOnse assessMent for oroPharyngeaL cancEr paTiEnts; Protocol for an observational study	[122]	2022	radiosensitivity	[3]
24	10.1016/j.radonc.2023.109868	Voxel-based analysis: Roadmap for clinical translation	[123]	2023	radiosensitivity	[3]
25	10.1002/cnr2.1126	Progressive breast fibrosis caused by extreme radiosensitivity: Oncocytogenetic diagnosis and treatment by reconstructive flap surgery.	[124]	2019	radiosensitivity, DNA repair	[3, 1]
26	10.1002/mp.15177	Radiobiological comparison between Cobalt-60 and Iridium-192 high-dose-rate brachytherapy sources: Part I—cervical cancer	[125]	2021	radiosensitivity, hypoxia-induced	[3, 5]
27	10.1016/j.ijrobp.2015.11.013	Influence of Nucleoshuttling of the ATM Protein in the Healthy Tissues Response to Radiation Therapy: Toward a Molecular Classification of Human Radiosensitivity	[126]	2016	radiosensitivity, radioresistance	[3, 4]
28	10.3390/proteomes13020025	Deciphering Radiotherapy Resistance: A Proteomic Perspective	[127]	2025	radiosensitivity, radioresistance, DNA repair, precision oncology	[3, 4, 1, 2]

^{* Classes: (1) DNA repair, (2) Precision oncology, (3) Radiosensitivity, (4) Radioresistance, and (5) Flash radiotherapy.}

By analyzing these elements together, the study made the physical foundations of the techniques more tangible through models and quantitative metrics, while the examination of device architectures and workflow constraints provided a realistic perspective on feasibility. Early clinical findings offered valuable reference points for assessing translational potential. In this way, the discussion remained coherent and comparative, reflecting the main thematic axes of oncological radiotherapy, both clinical and molecular, and showing how technological innovation acts as a connecting element that integrates these two dimensions within contemporary practice.

The questions used as subject, according to each thematic class, were as follows:

Classe 1 – DNA Repair and Molecular Response:What are the key molecular responses to radiation discussed in the document, including DNA damage signaling, DNA repair pathways, and checkpoint activation mechanisms?

Classe 2 – Precision Oncology and Genomic Modeling:How does the document address precision oncology, including the use of genomic profiling, machine learning models, and patient stratification in radiation therapy?

Classe 3 – Individual Radiosensitivity and Clinical Risk:What evidence does the document present on interindividual radiosensitivity, clinical risk assessment, and predictive biomarkers for radiation response?

Classe 4 – Radioresistance and Associated Mechanisms:What mechanisms of radioresistance are described in the document, including tumor hypoxia, metabolic reprogramming, stem cells, and viral integration?

Classe 5 – Advanced Technologies and Innovative Radiotherapy:How does the document explore advanced radiotherapy strategies, including FLASH, hadron therapy, voxel-based analysis, and dose enhancement with high-Z nanoparticles?

Cosine similarity is a measure of the extent to which two texts point in the same direction in feature space, and for non-negative representations (such as TF-IDF), it ranges from 0 (the texts share no terms and are therefore unrelated) to 1 (the texts have maximal lexical or semantic overlap). For the 37 text pairs compared here, cosine similarity scores ranged from 0.291 to 0.669 (span of ~0.378), with an overall mean of 0.5235 0.0271 (95% CI: 0.4964-0.5506), which indicates a moderate degree of similarity between the outputs of Llama3 (8B) and GPT-4o. This moderate degree of similarity is further broken down by class, where we find the following: Class 1 (n = 6): 0.4693 0.1542 (0.3151-0.6235), Class 2 (n = 6): 0.5387 0.0439 (0.4948-0.5826), Class 3 (n = 18): 0.5520 0.0289 (0.5231-0.5810), Class 4 (n = 4): 0.4592 0.0869 (0.3723-0.5460), and Class 5 (n = 3): 0.5167 0.2541 (0.2626-0.7708), and these classes show that Classes 2 and 3 have the highest means with the tightest confidence intervals, suggesting that Llama3 (8B) is able to reliably reproduce GPT-4o content in those domains, while Classes 1 and 4 have lower averages with wider intervals, indicating greater thematic or stylistic variability and sensitivity to domain-specific terminology, and Class 5 is inconclusive due to its very small sample size and wide uncertainty.

Cosine similarity was used as a robustness check only to compare lexical and semantic congruence between two LLMs, given the same prompts, because it measured congruence, but did not influence any bibliometric or clinical inferences. The mean similarity value of ~0.52 provides evidence for moderate cross-model reproducibility, and thereby, methodological robustness, and consequently, no hybrid inference or model fusion was performed; instead, the outputs were compared for completeness and terminological accuracy, prior to manual curation. The notation "Llama → GPT-4o" reflects only the editorial workflow (i.e., Llama3 generated the first draft and GPT-4o refined the language)without altering any analytic results, which improved clarity, consistency, and readability, while maintaining all quantitative and bibliometric content. While Llama3 remains a useful tool for fully offline analyses, which are important for data security and sovereignty, the comparison provided here demonstrates cross-model consistency, rather than hybrid modeling, and therefore serves to strengthen transparency and reproducibility in the pipeline, thus maintaining the integrity of the results.

3.1. CASE STUDIES

Class 1 studies (DNA Repair and Molecular Response) focus on the molecular mechanisms of radiation response, including DNA repair, checkpoint regulation, and therapeutic modulation. Norbert Mészáros et al. (2019) [124] reported progressive breast fibrosis in patients with extreme radiosensitivity, associated with high rates of chromosomal aberrations, supporting the use of cytogenetic tests for risk stratification. Luis Bermúdez-Guzmán et al. (2021) [102] demonstrated that chronoradiotherapy, by aligning dose delivery with the circadian rhythms of genes such as BMAL1, CLOCK, PER, and CRY, accelerates DNA break resolution and reduces toxicity. Michael D. Story et al. (2024) [101] showed that high-LET radiation induces complex lesions and activates the cGAS–STING pathway, with radiosensitization enhanced by inhibitors of HR, NHEJ, ATM, and ATR. Olga A. Martin et al. (2024) [103] reviewed 75 years of radiobiology, highlighting the impact of radiation quality on γH2AX persistence and the tumor microenvironment. Davide Perico and Pierluigi Mauri (2025) [127] correlated the overexpression of RAD51 and BRCA1 and the hyperactivation of NHEJ with radioresistance, while Yan Luo (2025) [104] demonstrated that high-Z nanoparticles amplify damage and reactive species, proposing standardization of metrics such as NER and SER.

Class 2 (Precision Oncology and Genomic Modeling). Ivana Dokic et al. (2016) [106] used ion beams and the Cell-Fit-HD technology to demonstrate that the persistence of γH2AX foci after 72h is a stronger biomarker of radiosensitivity than the initial lesion count, suggesting the development of biodosimetric repositories. Paolo Tini et al. (2018) [108], analyzing 17,412 cases of lung carcinoma, showed that radiotherapy provides the greatest benefit in advanced stages and squamous histology, advocating for the integration of molecular data. Henning Willers et al. (2019) [109] proposed the use of RSI and GARD to guide personalized dosing. Perico and Mauri (2025) [127] identified proteins such as RAD51, PARP1, CHK1, and MAPK15 as associated with resistance, while Marco Calvaruso et al. (2025) [107] argued for multi-omic biomarkers and artificial intelligence to predict treatment response. Yan Luo (2025) [104] developed a multidimensional index for high-Z nanoparticles, incorporating biological variability and immune modulation.

Class 3 (Individual Radiosensitivity and Clinical Risk) gathered the largest number of studies. Igor Sirák et al. (2015) [118] reported FANCA mutations associated with severe toxicities, while Adeline Granzotto et al. (2016) [126] identified ATM nucleoshuttling kinetics as a functional marker. Yasuhiro Ogawa (2016) [119] introduced KORTUC II as a radiosensitizer based on H₂O₂. Sarah L. Kerns et al. (2018) [120] reviewed the use of RSI and GARD, and E.J. Her et al. (2018) [121] modeled TCP for prostate brachytherapy. Nathalie Arians et al. (2019) [111] demonstrated that carbon ions overcome HPV-induced resistance. Mészáros et al. (2019) [124] correlated cytogenetic profiles with severe fibrosis. John Kang et al. (2020) [117] reviewed genomic and machine learning models, while Chekhun and Domina (2021) [114] suggested that COVID-19 may increase radiosensitivity. Dayyani et al. (2021) [125] compared ⁶⁰Co and ¹⁹²Ir in cervical brachytherapy, and Cesare Cozzarini et al. (2022) [112] developed a TCP model validated in 795 post-prostatectomy patients. Nikitović and Stanojković (2022) [116] linked microRNAs and cytokines to prostate cancer toxicity. Verduijn et al. (2022) [122] developed the COMPLETE protocol, integrating multiparametric imaging, omics data, and machine learning. Raitanen et al. (2023) [115] showed greater radioresistance in 3D spheroids, while McWilliam et al. (2023) [123] highlighted voxel-based analysis to map critical regions. Jazmati et al. (2025) [113] assessed pediatric medulloblastoma, identifying homogeneous α/β values. Perico and Mauri (2025) [127] emphasized proteomics for identifying radioresistance profiles, and Guo et al. (2025) [15] reviewed mechanisms of FLASH-RT, including oxygen depletion and mitochondrial preservation, highlighting its integration with immunotherapy.

Class 4 (Radioresistance and Associated Mechanisms). Granzotto et al. (2016) [126] linked delayed ATM nucleoshuttling to radioresistance. Mery et al. (2018) [110] showed that VHL mutations induce pseudohypoxia in renal carcinoma through HIF activation, suggesting carbon ions and repair inhibitors as therapeutic options. Arians et al. (2019) [111] demonstrated that carbon ions restore checkpoint control in HPV-positive cervical tumors with disrupted E2, with RBE values ranging from 1.3 to 4.3. Perico and Mauri (2025) [127] integrated pathways of hypoxia, metabolism, stemness, viral integration, and apoptosis evasion into proteomic maps, advocating their use for biomarker discovery and selective therapies.

Class 5 (Advanced Technologies and Innovative Radiotherapy) addresses the adoption of cutting-edge modalities. Zhang et al. (2021) [105] investigated FLASH-RT (≥40 Gy/s), showing comparable efficacy to conventional radiotherapy but with enhanced tissue preservation and immune modulation, including lymphocyte sparing and tumor remodeling with increased CD8⁺ infiltration. Dayyani et al. (2021) [125] compared ⁶⁰Co and ¹⁹²Ir in cervical cancer, concluding that both can be used with dose adjustments, with physical advantages of ⁶⁰Co and higher RBE for ¹⁹²Ir. Guo et al. (2025) [15] further explored the mechanisms of FLASH-RT, such as free radical modulation, mitochondrial preservation, and vascular integrity, emphasizing the need for dosimetric standardization and voxel-based mapping. Collectively, these studies consolidate the integration of FLASH-RT and optimized brachytherapy into next-generation personalized protocols.

The synthesis of the case studies presented above provides the scientific foundation for structuring the Integrated Implementation Plan in Precision Radiotherapy. Each class analyzed—from DNA molecular response to the adoption of innovative technologies such as FLASH-RT—revealed mechanisms, biomarkers, models, and strategies directly translatable into personalized clinical protocols. By organizing these findings into five interdependent operational blocks, the plan transforms experimental and translational evidence into a continuous flow of investigation, validation, and clinical application, establishing a framework capable of supporting both individual therapeutic advances and the incorporation of objective indicators into public health policies.

3.2. IMPLEMENTATION PLAN IN PRECISION RADIOTHERAPY

The Integrated Implementation Plan in Precision Radiotherapy (IIPPR) is organized into five classes that structure a continuous flow of investigation, validation, and clinical application, with the potential for direct integration into public health policies. Class 1 – DNA Repair and Molecular Response constitutes the operational block responsible for establishing the molecular and functional foundation upon which all subsequent stages are built. Its objective is to identify DNA repair mechanisms, regulate checkpoints, and characterize signaling pathways that determine each patient’s response to radiation. This step provides indispensable biomarkers and functional parameters for therapeutic personalization, enabling biologically precise decisions regarding technique, dose, fractionation, and adjuvant combinations from the outset of the workflow. Implementation requires functional assays, omics analyses, and molecular modeling at multiple levels of complexity.

Norbert Mészáros et al. (2019) [124] reported that patients with progressive radiation-induced breast fibrosis exhibited high frequencies of chromosomal aberrations even in non-irradiated cells, demonstrating that genomic instability and persistent checkpoint failures can be detected cytogenetically and used as predictive tools. Temporal modulation strategies, as shown by Luis Bermúdez-Guzmán et al. (2021) [102], revealed that synchronizing irradiation with circadian phases of maximal repair activity accelerates the resolution of double-strand breaks and enhances the efficiency of HR, NHEJ, NER, and BER, reducing toxicity and improving efficacy. Michael D. Story et al. (2024) [101] demonstrated that high-LET radiation, such as that used in hadron therapy, induces complex lesions and activates immune pathways via cGAS–STING, whose radiosensitizing potential can be amplified by pharmacological inhibition of HR, NHEJ, ATM, and ATR. The review by Olga A. Martin et al. (2024) [103] consolidated the relevance of radiation quality in the persistence of γH2AX, remodeling of the tumor microenvironment, and integration with immunotherapy. Davide Perico and Pierluigi Mauri (2025) [127] highlighted proteomic integration as a tool to identify functional profiles of key proteins associated with radioresistance, while Yan Luo (2025) [104] demonstrated that high-Z nanoparticles enhance radiosensitization by overloading repair pathways, with standardization in metrics such as NER and SER.

The central action of this class is to integrate different levels of molecular and functional information into a validated panel of predictive biomarkers, delivering to subsequent phases a comprehensive map of tumor vulnerabilities and the repair limitations of normal tissues.

Class 2 – Precision Oncology and Genomic Modeling transforms molecular and radiobiological data into parameters guiding the choice of technique, dose, and fractionation, replacing empirical protocols with personalized prescriptions. Ivana Dokic et al. (2016) [106], using Cell-Fit-HD technology, showed that repair kinetics is a more robust marker than initial lesion count and proposed biodosimetric repositories to support biological prescriptions. In a population-based analysis, Paolo Tini et al. (2018) [108] examined 17,412 cases of non-small cell lung carcinoma and demonstrated the need to integrate genomic markers and predictive models to improve patient selection. Henning Willers et al. (2019) [109] proposed indices such as RSI and GARD to calibrate dose based on gene expression, dosimetry, and imaging, further integrating immunotherapy and real-time adjustments. Davide Perico and Pierluigi Mauri (2025) [127] added functional proteomic layers to stratification, identifying proteins such as RAD51, PARP1, CHK1, and MAPK15, thereby enhancing predictive power when combined with transcriptomic data. Marco Calvaruso et al. (2025) [107] demonstrated that multi-omic biomarkers, combined with FLASH-RT and predictive algorithms, enable fine-tuning of intensity and technique while minimizing toxicity. Yan Luo (2025) [104] developed a multidimensional index integrating biological variability, immune modulation, and DER, SER, and RBE metrics. Recent advances in endometrial cancer research highlight the role of genomic modeling in improving radiotherapy through molecular classification. The PORTEC-3 study showed that distinguishing POLE-mutated from TP53-mutated tumors predicts treatment response and prognosis, supporting adaptive dosing based on molecular phenotype. The PORTEC-4 trial is expanding this approach by integrating molecular data into therapeutic decision-making, aiming for biologically guided radiotherapy optimization. The TCGA classification defines four subtypes, POLE-mutated, MMRd, p53abn, and NSMP, each with distinct prognostic value. POLE and MMRd tumors show favorable outcomes, while p53abn tumors are linked to poor prognosis [128,129,130,131,132,133,134,135,136,137]. Incorporating these molecular markers into clinical workflows enables more precise, personalized radiotherapy, improving efficacy and patient outcomes. This class delivers a therapeutic plan calibrated to both tumor and patient, pre-tested in predictive simulations and ready for clinical application.

Class 3 – Individual Radiosensitivity and Clinical Risk is the block dedicated to identifying and integrating variables that define individual tolerance to radiation, adjusting therapeutic protocols according to genetic, functional, and clinical predispositions. Igor Sirák et al. (2015) [118] documented exacerbated toxicity in a patient with pathogenic heterozygosity in FANCA, while Adeline Granzotto et al. (2016) [126] showed that delayed nuclear translocation of ATM compromises repair—both examples of biomarkers for prior screening. Yasuhiro Ogawa (2016) [119] introduced the KORTUC II method, combining H₂O₂ modulation, and Sarah L. Kerns et al. (2018) [120] and John Kang et al. (2020) [117] explored RSI and GARD in association with TCP and NTCP models. E.J. Her et al. (2018) [121] highlighted the influence of α-parameter variability in brachytherapy, and Nathalie Arians et al. (2019) [111] demonstrated that carbon ions overcome HPV-induced resistance. Norbert Mészáros et al. (2019) [124] reinforced the value of cytogenetics, and V.F. Chekhun & E.A. Domina (2023) [114] expanded the discussion to systemic effects such as those associated with COVID-19. Mahdieh Dayyani et al. (2021) [125] compared ⁶⁰Co and ¹⁹²Ir sources in brachytherapy, while Cesare Cozzarini et al. (2022) [112] developed TCP models validated in post-prostatectomy cohorts. Marina Nikitović & Tatjana Stanojković (2022) [116] compiled molecular and clinical evidence, while Gerda M. Verduijn et al. (2022) [122] consolidated the COMPLETE protocol, integrating multiparametric imaging, radiomics, and machine learning. Recent studies expanded the translational basis with 3D spheroids (Raitanen et al., 2023) [115], voxel-based mapping (McWilliam et al., 2023) [123], pediatric radiobiology (Jazmati et al., 2025) [113], functional proteomics (Perico & Mauri, 2025) [127], and FLASH-RT (Guo et al., 2025) [15]. The central action of this class is to consolidate an individual risk matrix combining biomarkers and clinical data, transforming them into operational recommendations for dose, technique, and modality.

Class 4 – Radioresistance and Associated Mechanisms ocuses on identifying and neutralizing tumor mechanisms of resistance to radiation. Adeline Granzotto et al. (2016) [126] showed that delayed ATM phosphorylation kinetics compromises target activation, while Mery et al. (2018) [110] highlighted the role of hypoxia and VHL mutations in renal carcinoma, activating pro-survival pathways. Nathalie Arians et al. (2019) [111] demonstrated that in HPV-positive cervical cancer, the loss of cell-cycle control and degradation of p53 and Rb can be reversed through carbon-ion irradiation. Davide Perico and Pierluigi Mauri (2025) [127] expanded this framework through high-resolution proteomics, integrating axes of hypoxia, metabolism, stemness, viral integration, and enhanced repair. This class translates mapping into interventions such as high-LET modalities, repair inhibitors, and microenvironment modulators, effectively reversing resistant phenotypes and improving outcomes in poor-prognosis subgroups.

Class 5 – Advanced Technologies and Innovative Radiotherapy iis dedicated to the incorporation of next-generation modalities. Zhang et al. (2021) [105] demonstrated that ultrafast radiotherapy (FLASH-RT, ≥40 Gy/s) maintains tumor efficacy while preserving normal tissues and remodeling the immune microenvironment. Guo et al. (2025) [15] reinforced the clinical feasibility of FLASH-RT, identifying multiple protective mechanisms—including mitochondrial preservation and CD8⁺ lymphocyte maintenance—provided it is accompanied by dosimetric standardization. Mahdieh Dayyani et al. (2021) [125] showed that the comparison of ⁶⁰Co and ¹⁹²Ir in brachytherapy allows for strategic adjustments, balancing biological efficacy and safety. This class integrates FLASH-RT, optimized brachytherapy, and voxel-based mapping into adaptive protocols that combine physical precision with biological selectivity.

The five classes together describe a pragmatic implementation plan for connecting precision radiotherapy to population-level public health delivery, and for Brazil's Unified Health System (SUS), the plan describes how standardized workflows, such as molecular diagnostics, biomarker-driven patient stratification, and adaptive radiotherapy protocols, can be implemented in a step-wise fashion, if and when clinically validated. Rather than promise immediate gains, the framework highlights the essential interfaces between biomarker discovery, dosimetric modeling, and outcome assessment, because once validated in pilot environments, these interfaces will enable more evidence-based decisions and rational distribution of limited resources.

Table 2 organizes these prospective metrics by class, describing how survival, toxicity, personalization, access, and cost-effectiveness indicators might be measured in pilot or registry-based studies. Figure 2 presents a conceptual integration map that connects the plan’s classes (blue nodes), measurable indicators (green nodes), and strategic health objectives (red nodes), illustrating a logical pathway from scientific discovery to policy evaluation rather than a proven causal link. This structure positions the plan as a dynamic and testable framework that links molecular biology, clinical radiotherapy, and health management in a coherent model open to validation through prospective cohorts and pilot programs within the SUS.

3.3. Contributions, Practical Implications, Limitations, and Future Work

This paper provides a timely, simple, and humanistic account of how radiotherapy, radiobiology, and oncology have evolved and are evolving, based on the AI-based semantic and temporal analysis of 3,343 publications between 1964 and 2025. The most significant strength of this paper lies in its methodological integration: rigorous data cleaning and normalization (deduplication and lemmatization), powerful discovery tools (topic modeling and network analysis), and careful temporal mapping to reconstruct the conceptual architecture of the field, which, from this perspective, identifies two complementary axes that presently define the field.

The clinical-anatomical axis classifies disease sites, therapeutic modalities, and prognostic endpoints. In contrast, the mechanistic-molecular axis describes DNA damage and repair, biomarkers, and cellular radiation responses, and together, they describe a field that has matured from a collection of isolated silos into an interdependent system, where biological mechanisms and clinical decision-making inform one another in near real time.

The most practical contribution of this paper is the Precision Radiotherapy Implementation Plan, which operationalizes this integrated understanding into a practical framework, defining five classes - DNA repair and molecular response, precision oncology and genomic modeling, individual radiosensitivity and clinical risk, mechanisms of radioresistance, and innovative radiotherapy technologies - each integrating biomarkers, predictive models, and therapeutic strategies in support of individualized care.

Importantly, this is not a conceptual roadmap, but a practical bridge from data to bedside, structured to align AI-derived biological markers with real-world outcomes that matter to patients and healthcare systems: overall and progression-free survival, toxicity profile, and cost per controlled case, because, in the context of a scalability- and equity-oriented environment like Brazil's SUS, the framework describes how to integrate molecular diagnostics and predictive stratification in resource-sensitive, standardized pathways. It also describes a feasible path to the responsible implementation of high-end modalities such as FLASH-RT and hadron therapy by relating their deployment to measurable benefits and real-world feasibility.

The authors are forthright about the limitations of the study, because the dip in 2025, for instance, represents an indexing lag rather than an actual decline in research output. The authors also confined the primary analysis to 2014-2025 and used a controlled set of 15 standardised terms to keep the meanings consistent between sources, which improves internal consistency and semantic precision, but also means that the analysis captures less of the earlier historical record.

The deliberate exclusion of non-DOI records and pre-2014 literature, while improving data quality and comparability, may have omitted some historically important or recently under-indexed contributions. Nevertheless, this methodological choice has produced a clean, reusable baseline that can be expanded in future studies to encompass longer temporal horizons.

Some thematic clusters—particularly Advanced Technologies—suffered from limited sample sizes, resulting in wider uncertainty bands and validation challenges. Additionally, heterogeneous terminology and writing styles across sources posed difficulties for semantic alignment, especially when processed with models such as Llama-3 (8B). These challenges are intrinsic to large-scale, multilingual, and rapidly evolving biomedical corpora, and they define the natural constraints of first-generation integrative analyses.

Looking forward, the authors outline a clear roadmap for methodological refinement. First, coupling physical and biological models of ultrafast radiotherapy with high-Z nanotechnologies could enhance system-level dynamic modeling. Second, semantic supergraph analysis may uncover underexplored clusters, early signals, and strategic opportunities for funding and collaboration. Third, incorporating AI-derived biomarkers into adaptive, real-time radiotherapy planning will enable predictive systems that evolve with patient response, closing the loop between inference and action.

This work represents a milestone in computational oncology. It demonstrates that integrative synthesis in radiotherapy is not only feasible but transformative when paired with clinical expertise and public-health pragmatism. By weaving heterogeneous data into a cohesive, decision-ready framework, the study advances the field toward a continuously learning ecosystem where semantic analytics, molecular insights, and clinical innovation converge to deliver more precise, safer, and equitable cancer care.

4. Discussion

These analyses detail how modern radiotherapy has evolved historically, semantically and operationally into an interdisciplinary, data-driven ecosystem spanning radiobiology and oncology, because the computational workflow began with automated retrieval from Scopus, PubMed and Web of Science, followed by standardization, normalization and metadata integration, resulting in 3,343 unique records. The source imbalance, with PubMed as the major contributor, illustrates the field's biomedical origins, and the temporal profile of publication activity shows three phases: a low-output phase (1964-1990), steady growth phase (1991-2010) and high-output phase (2011 onwards) coinciding with the development of stereotactic body radiotherapy, proton therapy and predictive modeling, although the apparent decline in 2025 is best explained by indexing delay rather than a real decrease in output. A word cloud of titles and abstracts displays recurrent terms-cell, cancer, dose, treatment, radiotherapy-underscoring the continued interplay between mechanistic biology, dosimetry and clinical practice, and Latent Dirichlet Allocation revealed ten major topics across titles and abstracts, categorized into two interrelated axes, namely a clinical-anatomical axis containing tumor sites, treatment modalities and patient-centered outcomes, and a mechanistic-molecular axis including cellular responses, DNA repair, biomarkers, and toxicity pathways. Together, these axes form a translational continuum where experimental findings and clinical needs are informing each other iteratively, and the network topology of the co-occurrence supergraph has a high density (0.07), low modularity (0.15), and clustering coefficient (0.5), suggesting a strong semantic cohesion, because the central node cancer is directly connected to tumor, radiotherapy, particle, and breast, while the technological subfields (proton, stereotactic, and imaging) that anchor the methodological innovation are positioned at the periphery. The findings reveal a deep conceptual convergence across biology, medical physics, and clinical care in radiotherapy, marking its transition from siloed subdisciplines to an integrated translational science, and this model supports the following four working hypotheses: cancer and radiotherapy are the structural axes of recent radiotherapy scholarship, second, the increasing frequencies of cell, dose, effect, and treatment reflect the increasing focus on personalization and clinical efficacy, third, the cooccurrence of tumor, DNA, repair, and survival suggests that precision-medicine research has increased, and fourth, high density and low modularity reflect the transversal and integrative character of the field. Targeted database searches based on these hypotheses identified 61 articles of high relevance published between 2014 and 2025, which were grouped into five themes: DNA repair and molecular responses, precision oncology and genomic modeling, individual radiosensitivity and clinical risk, mechanisms of radioresistance, and advanced technologies and innovation in radiotherapy, and case analyses demonstrate how these themes complement each other, because DNA repair studies focused on complex damage from high-LET radiation, circadian alignment of therapy, and the modulation of the HR, NHEJ, ATM, and ATR pathways, while precision oncology studies integrated multi-omic data and AI models for dose calibration and genomic stratification, using indices such as RSI and GARD, and protein markers such as RAD51, PARP1, CHK1, and MAPK15. Individual radiosensitivity studies combined genetic and clinical data to develop personalized risk matrices based on TCP/NTCP modeling, and including imaging, radiomics, and machine learning, and the radioresistance studies highlighted the potential of high-LET modalities, repair inhibitors and proteomic mapping to reverse resistant phenotypes, whereas the technology studies reviewed the potential of FLASH radiotherapy and optimized brachytherapy for tissue-sparing effects and immune modulation. These studies provided the foundation for the Integrated Implementation Plan in Precision Radiotherapy, an operational framework of five interdependent blocks that connects molecular discovery to translational validation and clinical delivery, because it utilizes biomarker panels, genomic models and adaptive protocols to support biologically informed treatment planning, and is designed with Brazil's Unified Health System in mind, specifying measurable indicators-overall and progression-free survival, local control, grade ≥3 toxicity, unplanned hospitalizations and average cost per controlled case-that align with health-system goals to improve equity, reduce mortality, optimize resources and expand access to advanced technologies. Together, they form a practical, evidence-based pathway for oncology, and conceptually and methodologically, this study demonstrates that the combination of semantic mining, temporal modeling and network analysis can reconstruct the evolution of a complex biomedical domain, therefore radiotherapy is shown to be a coherent, interdependent system in which molecular biology, medical physics and clinical decision-making reinforce each other in a continuous learning loop. However, the authors note several limitations, including the likely index lag in 2025, the choice of analytic window (2014-2025) and the exclusion of non-DOI records, although these choices enhanced consistency and semantic precision at the expense of historical depth, heterogeneous terminology and limited published data for certain advanced technologies likely expanded the confidence intervals of certain results, and nevertheless, the Pearson values (>0.9) indicate the robustness of the main interpretations. Consequently, the implications of this study are substantial, because theoretically, the work provides a fine-grained conceptual map of modern radiotherapy, operationally, it provides a reproducible framework for the implementation of personalized, data-guided care pathways, and for health policy, it outlines a scalable model for the integration of advanced technologies within public healthcare systems, thus future directions include the development of coupled physical and biological models of ultrafast radiotherapy coupled with high-Z nanoparticle platforms, expansion of semantic supergraph analysis to surface underexplored clusters and embedding AI-derived biomarkers into adaptive, real-time learning protocols. Moreover, the confluence of semantic analytics, molecular modeling and clinical innovation described above leads to an ever learning oncology ecosystem that transforms data into safer, more precise and more equitable cancer care.

5. Conclusions

The current study demonstrates that AI-assisted semantic and temporal analysis of the scientific literature in radiotherapy, radiobiology, and oncology provides a sound and forward-looking foundation for a precision implementation agenda that bridges molecular science, clinical practice, and public health, because by analyzing 3,343 unique documents published between 1964 and 2025, the study reconstructs the historical trajectory and conceptual architecture of the field, showing how radiotherapy has transitioned from a fragmented domain to an integrated, interdisciplinary system, characterized by semantic cohesion and translational continuity.

The results delineate a mature scientific landscape located at the intersection of biology, physics, and clinical oncology, with strong semantic linkages among core concepts including cancer, radiotherapy, DNA repair, radiation dose, and patient outcomes, and two dominant research dimensions are identified: the first is clinical-anatomical, centered on tumor classification, therapeutic modalities, and prognosis, while the second is mechanistic-molecular, focused on DNA repair pathways, biomarkers, and cellular responses to radiation.

The convergence of these dimensions highlights the translational character of the field, in which molecular findings are rapidly operationalized into therapeutic and technological innovations with direct patient impact; therefore, the study outlines the Integrated Implementation Plan in Precision Radiotherapy, a structured model that operationalizes scientific evidence into an iterative cycle of discovery, validation, and clinical deployment.

The programme is structured around five operational classes: DNA repair and molecular response; precision oncology and genomic modelling; individual radiosensitivity and clinical risk; mechanisms of radioresistance; and innovative radiotherapy technologies, and each class incorporates biomarkers, predictive modelling, and therapeutic strategies to allow the development of personalized treatment plans with the overall objective of improving survival, reducing toxicity, and increasing equitable access to advanced technologies including FLASH-RT and hadron therapy.

The programme addresses the perennial problem of data silos in biomedical research by harmonizing heterogeneous data sets using AI-based modelling, thereby turning isolated findings

into a coherent, decision-ready evidence base, thus placing AI at the heart of interpreting and integrating the complexity of modern science.

Overall, the programme suggests that radiotherapy, radiobiology, and oncology have reached a high level of interdisciplinary maturity that now supports the development of a dynamic and integrated framework for precision medicine, moreover, in addition to computational synthesis, the programme represents a conceptual advance in the organization and application of scientific knowledge allowing for decades of data transformed into a meaningful base for innovation, clinical decision-making, and more equitable care.

Appendix A.1. Topics from Titles (t)

Topic_0t = 0.090*"cancer" + 0.050*"radiotherapy" + 0.049*"patient" + 0.038*"head" + 0.037*"neck" + 0.022*"breast" + 0.018*"prostate" + 0.013*"toxicity" + 0.011*"advanced" + 0.010*"treatment" Eq. 1

Topic_1t = 0.075*"radiobiology" + 0.040*"radiotherapy" + 0.036*"clinical" + 0.031*"radiation" + 0.026*"therapy" + 0.019*"radiobiological" + 0.016*"treatment" + 0.015*"oncology" + 0.013*"perspective" + 0.010*"new" Eq. 2

Topic_2t = 0.050*"cancer" + 0.042*"radiotherapy" + 0.021*"patient" + 0.021*"breast" + 0.018*"carcinoma" + 0.014*"treatment" + 0.013*"prostate" + 0.012*"cell" + 0.010*"esophageal" + 0.009*"comparison" Eq. 3

Topic_3t = 0.097*"radiation" + 0.046*"oncology" + 0.023*"therapy" + 0.022*"biology" + 0.020*"mechanism" + 0.014*"molecular" + 0.011*"medical" + 0.011*"injury" + 0.011*"effect" + 0.011*"clinical" Eq. 4

Topic_4t = 0.058*"stereotactic" + 0.033*"body" + 0.031*"therapy" + 0.029*"lung" + 0.028*"cancer" + 0.028*"radiotherapy" + 0.027*"cell" + 0.025*"radiosurgery" + 0.023*"tumor" + 0.021*"radiation" Eq. 5

Topic_5t = 0.044*"radiation" + 0.042*"cancer" + 0.040*"therapy" + 0.024*"cell" + 0.021*"tissue" + 0.020*"normal" + 0.018*"beam" + 0.018*"ion" + 0.013*"proton" + 0.013*"effect" Eq. 6

Topic_6t = 0.039*"radiotherapy" + 0.029*"tumor" + 0.020*"cancer" + 0.016*"radiation" + 0.015*"therapy" + 0.015*"patient" + 0.015*"brain" + 0.012*"model" + 0.011*"imaging" + 0.011*"preclinical" Eq. 7

Topic_7t = 0.049*"tumor" + 0.034*"cell" + 0.026*"dna" + 0.023*"repair" + 0.022*"factor" + 0.019*"cancer" + 0.014*"lung" + 0.013*"targeting" + 0.013*"radiotherapy" + 0.013*"pathway" Eq. 8

Topic_8t = 0.093*"cell" + 0.039*"human" + 0.028*"effect" + 0.023*"expression" + 0.022*"radiation" + 0.018*"gene" + 0.017*"carcinoma" + 0.016*"line" + 0.015*"vitro" + 0.014*"irradiation" Eq. 9

Topic_9t = 0.089*"dose" + 0.033*"radiotherapy" + 0.030*"rate" + 0.029*"brachytherapy" + 0.018*"high" + 0.016*"radiobiology" + 0.016*"prostate" + 0.015*"model" + 0.013*"linear" + 0.012*"radiobiological" Eq. 10

Appendix A.2. Topics from Abstracts (a)

Topic_0a = 0.033*"model" + 0.013*"dose" + 0.012*"tumor" + 0.012*"imaging" + 0.009*"flash" + 0.009*"feature" + 0.008*"clinical" + 0.008*"image" + 0.007*"parameter" + 0.006*"volume" Eq. 11

Topic_1a = 0.070*"cell" + 0.021*"tumor" + 0.014*"effect" + 0.013*"dna" + 0.009*"repair" + 0.009*"response" + 0.009*"damage" + 0.008*"line" + 0.007*"mechanism" + 0.006*"survival" Eq. 12

Topic_2a = 0.033*"proton" + 0.025*"ion" + 0.022*"rbe" + 0.021*"beam" + 0.016*"particle" + 0.014*"carbon" + 0.014*"energy" + 0.014*"biological" + 0.013*"therapy" + 0.012*"photon" Eq. 13

Topic_3a = 0.044*"patient" + 0.016*"rt" + 0.015*"survival" + 0.015*"tumor" + 0.011*"p" + 0.008*"surgery" + 0.008*"local" + 0.008*"rate" + 0.008*"overall" + 0.007*"o" Eq. 14

Topic_4a = 0.056*"patient" + 0.019*"toxicity" + 0.015*"p" + 0.011*"breast" + 0.010*"risk" + 0.010*"grade" + 0.010*"gy" + 0.010*"month" + 0.008*"sbrt" + 0.008*"treated" Eq. 15

Topic_5a = 0.033*"study" + 0.028*"therapy" + 0.022*"clinical" + 0.017*"trial" + 0.015*"preclinical" + 0.012*"tumor" + 0.012*"immunotherapy" + 0.011*"targeted" + 0.010*"agent" + 0.010*"promising" Eq. 16

Topic_6a = 0.023*"expression" + 0.019*"cell" + 0.017*"gene" + 0.012*"response" + 0.012*"patient" + 0.011*"tumor" + 0.008*"level" + 0.007*"irradiation" + 0.007*"protein" + 0.006*"blood" Eq. 17

Topic_7a = 0.026*"dose" + 0.023*"gy" + 0.020*"irradiation" + 0.014*"cell" + 0.013*"tumor" + 0.012*"plan" + 0.011*"dos" + 0.009*"day" + 0.009*"target" + 0.008*"mouse" Eq. 18

Topic_8a = 0.016*"clinical" + 0.012*"therapy" + 0.012*"oncology" + 0.008*"radiobiology" + 0.007*"patient" + 0.006*"right" + 0.006*"reserved" + 0.006*"medical" + 0.006*"development" + 0.006*"elsevier" Eq. 19

Topic_9a = 0.039*"dose" + 0.020*"tumor" + 0.014*"fraction" + 0.013*"tissue" + 0.012*"model" + 0.012*"effect" + 0.011*"gy" + 0.009*"normal" + 0.008*"rate" + 0.008*"cell" Eq. 20