INTRODUCTION
Malignant tumors continue to threaten human health because of their high incidence and mortality, imposing a substantial burden on socioeconomic development and healthcare systems[1]. Although continuous advances in surgery, radiotherapy, chemotherapy, targeted therapy, and immunotherapy have significantly improved patient survival, the marked heterogeneity and dynamic evolution of cancer still limit the effectiveness of conventional diagnostic and therapeutic models. These limitations have driven oncology toward precision medicine paradigms based on molecular subtyping, biomarker identification, and individualized risk assessment[2]. With the rapid development of high-throughput sequencing, digital pathology, medical imaging, and liquid biopsy technologies, cancer care has entered an era of multimodal data-driven decision-making[3]. However, the massive scale, heterogeneity, and high dimensionality of medical data also place greater demands on information extraction, feature integration, and clinical decision support. Artificial intelligence (AI), with its strengths in pattern recognition, feature learning, and predictive modeling, has been widely applied across multiple stages of cancer diagnosis and treatment[4,5]. In particular, AI has enabled the joint analysis of multi-omics, radiomics, and clinical information, thereby facilitating the transition of precision oncology from single-indicator judgment to multidimensional comprehensive assessment. As an important component of comprehensive cancer care, traditional Chinese medicine (TCM) has distinct advantages in alleviating clinical symptoms, reducing adverse reactions, improving quality of life, and promoting rehabilitation[6,7]. Its holistic perspective and treatment based on syndrome differentiation are theoretically compatible with modern concepts of whole-course cancer management and individualized intervention. Nevertheless, TCM research has long faced challenges, including the subjectivity of syndrome differentiation, high heterogeneity in clinical data, and the complexity of the mechanisms of Chinese herbs, all of which constrain its standardization, evidence-based development, and internationalization[8]. AI offers new opportunities to address these bottlenecks. On the one hand, AI can facilitate the structured organization and knowledge discovery of TCM classical texts, clinical records, and empirical expertise[9]; on the other hand, it can mine prescription compatibility patterns, screen active constituents, predict therapeutic targets, and model syndrome differentiation, thereby enhancing the objectivity, interpretability, and translational efficiency of TCM research[10,11]. Accordingly, AI-driven integrative research combining TCM and Western medicine represents a promising direction for improving the precision and efficacy of cancer diagnosis and treatment while advancing the modernization of TCM.
TECHNOLOGICAL FOUNDATIONS AND CLINICAL TRANSLATION OF PRECISION ONCOLOGY
Precision oncology is a diagnostic and therapeutic paradigm that is based on individual biological characteristics, tumor molecular subtypes, and clinical heterogeneity. It integrates multi-omics testing, medical imaging, pathological information, and dynamic follow-up data to formulate individualized strategies for prevention, diagnosis, treatment, and prognostic management[12]. Unlike conventional empirically driven practice based primarily on anatomical location and histological classification, precision oncology focuses on the molecular mechanisms of tumorigenesis and interpatient variability. It emphasizes stratified diagnosis, precise intervention, and continuous management through biomarker identification, risk stratification, and dynamic monitoring. With the rapid advancement of medical AI, the technical framework of precision oncology has been continuously refined and is gradually progressing from scientific exploration to clinical translation.
Key technical systems of precision oncology
From a methodological perspective, ML, DL, natural language processing (NLP), knowledge graphs (KGs), and graph neural networks (GNNs) together constitute the core technical foundation of AI-enabled cancer research. ML primarily addresses the analysis of structured data and can identify latent patterns in clinical records, laboratory indicators, and multi-omics features; it has therefore been widely applied to disease classification, patient stratification, prognostic assessment, and treatment response prediction[13]. DL exhibits stronger representation learning capabilities for unstructured data and is particularly suited to complex data types such as medical imaging, digital pathology, and clinical text. Through multilayer neural networks, DL can automatically learn high-order features and substantially reduce dependence on manual feature engineering[14]. NLP provides technical support for the structured extraction and in-depth utilization of textual data from electronic health records (EHRs), pathology reports, and biomedical literature, thereby greatly improving the analyzability and reusability of real-world data[15]. On this basis, KGs provide an important framework for modeling cancer-related biomedical knowledge by structurally representing and systematically integrating entities such as genes, proteins, diseases, drugs, and clinical phenotypes, together with their complex interrelationships[16]. GNNs further advance representation learning and relational inference in graph-structured data, enabling the exploration of multilevel biological mechanisms underlying tumorigenesis, progression, and therapeutic response from a network perspective[17]. In addition, multimodal learning synergistically integrates heterogeneous information from molecular, imaging, pathological, and clinical domains to construct more comprehensive patient representations, thereby providing stronger support for precise stratification, prognostic prediction, and individualized intervention[18]. Through these technological approaches, AI has shown substantial application value across multiple critical stages of precision oncology.
AI-driven integrated TCM and Western medicine oncology model architecture
Lin et al[19] proposed a multimodal fusion framework for intelligent TCM syndrome differentiation, integrating omics, imaging, and four-diagnosis data. It uses NLP and object detection for feature extraction and unified representation of multi-source heterogeneous data, with syndrome differentiation as the anchor. Knowledge graphs and large models perform semantic dimensionality reduction to map high-dimensional data precisely to the syndrome space, suppressing noise interference. For engineering deployment, serverless parallel stream computing and edge quantization enable efficient lightweight model implementation. Experiments show that, following TCM holism, the framework enhances accuracy and generalization via cross-modal complementarity, balancing algorithmic precision and clinical implementation (Figure 1).
Figure 1 Technical framework of artificial intelligence-driven multimodal data integration for integrated traditional Chinese medicine and Western medicine in precision oncology.
TCM: Traditional Chinese Medicine. Created by Figdraw (ID: RSYTT62959; Supplementary material).
Limitations and implementation challenges of AI in clinical oncology translation
It must be acknowledged that, constrained by algorithmic limitations, implementation risks, and the requirements for clinical deployment, AI has yet to achieve scalable clinical application in oncology, and significant shortcomings persist in its clinical translation.
In terms of interpretability, the current mainstream post-hoc explanation techniques have inherent limitations, and the industry has not yet established standardized evaluation criteria, resulting in the widespread problem of “black box” decision-making in AI models. The inability to trace the decision logic of models makes it difficult for clinicians to accurately judge the rationality of algorithm outputs, thereby giving rise to either excessive reliance or one-sided rejection of algorithm conclusions. This not only undermines the professionalism and credibility of multidisciplinary tumor diagnosis and treatment but also triggers compliance challenges concerning informed consent for diagnosis and treatment, as well as the division of medical liabilities, greatly hindering the safe implementation of AI in the clinical diagnosis and treatment of cancer.
In terms of model performance, the efficacy of AI algorithms is highly dependent on training datasets. When confronted with complex and dynamic real-world clinical scenarios, such algorithms are prone to domain shift with insufficient stability, which constitutes the core cause of algorithm failure. Meanwhile, data bias runs through the entire process of model training, validation, and deployment, covering multiple types, including historical statistics bias, sample sampling bias, data annotation bias, and omitted variable bias. Such biases accumulate and amplify layer by layer, inducing a variety of clinical risks. In the context of oncology diagnosis and treatment, data bias may lead to mismatched therapeutic regimens and distorted prognostic evaluation results, further resulting in overtreatment, undertreatment, and misdiagnosis of patients’ conditions. Additionally, it widens the diagnosis and treatment gap among different populations and impairs the fairness of oncology medical services.
From the perspective of research and implementation, most existing oncology AI models focus primarily on technological research and development without the in-depth and full participation of clinical experts throughout the whole process. This leads to disconnection between algorithm research and practical clinical needs, poor adaptability to complex clinical scenarios, and a substantial reduction in model reliability in high-risk decision-making for oncological diagnosis and treatment[20,21].
In conclusion, although AI has a broad application prospect in the field of precise tumor diagnosis and treatment, it is currently still in the initial exploration and pilot application stage. Only by effectively addressing core bottlenecks such as limited high-quality data, insufficient stability of algorithm generalization, lack of model interpretability, and an underdeveloped collaborative R&D mechanism among industry, academia, and research institutions, can AI achieve large-scale clinical implementation and develop into a safe, controllable, standardized, and widely applicable auxiliary tool for tumor diagnosis and treatment.
MODERNIZATION OF TRADITIONAL CHINESE MEDICINE IN CANCER PREVENTION AND TREATMENT AND THE INTEGRATED DIAGNOSTIC AND THERAPEUTIC PATHWAY OF TRADITIONAL CHINESE MEDICINE AND WESTERN MEDICINE
TCM has profound theoretical foundations and extensive clinical experience in cancer prevention and treatment. Its core principles are rooted in a holistic view, treatment based on syndrome differentiation, and preventive intervention before disease onset. According to TCM theory, the occurrence and progression of tumors do not merely reflect localized lesions but rather result from the long-term interaction of multiple factors, including deficiency of vital qi, dysfunction of the zang-fu organs, and pathological changes involving qi, blood, phlegm, blood stasis, and toxins[22]. Accordingly, the role of TCM in cancer prevention and treatment includes not only intervention against tumor progression but also improvement of the host’s internal environment, regulation of immune status, alleviation of treatment-related toxicities, delay of recurrence and metastasis, and enhancement of quality of life. This therapeutic approach, characterized by the integration of disease diagnosis and syndrome differentiation as well as holistic regulation, is to some extent compatible with modern oncology concepts concerning tumor heterogeneity, host-tumor interactions, the tumor microenvironment, and whole-course management[23,24]. However, the integrated diagnostic and therapeutic approach of TCM and Western medicine in oncology is not merely a simple combination of two therapeutic systems. Rather, it involves complementary strategies tailored to different disease stages, therapeutic goals, and patient conditions. In this context, AI can be used not only to improve recognition and prediction in individual tasks but also to serve as a crucial technical bridge for collaborative precision diagnosis and treatment between TCM and Western medicine. Its specific application value is mainly reflected in the following aspects.
Data mining: Construction of collaborative KGs for TCM and Western medicine
By harnessing multimodal fusion and intelligent algorithms, this approach helps decipher the intricate interplay between ancient textual connotations, clinical expertise, and molecular mechanisms. This approach effectively bridges the critical gaps of data fragmentation, limited dimensionality, and a lack of quantifiable metrics in oncological TCM research.
Data alignment: Multi-dimensional heterogeneous data fusion to build a knowledge base system
Data alignment is the foundation of intelligent analysis of TCM for oncology, aiming to integrate multi-source heterogeneous data such as ancient medical classics, clinical medical cases, drug targets, and disease molecules, and achieve the matching and correlation between macroscopic syndromes and microscopic biological mechanisms. Kang et al[25] based on the national medical classics database, systematically retrieved ancient Chinese medicine tumor-related disease names, and through automated data acquisition and visual analysis, sorted out their evolution history and development overview, providing a methodological paradigm for the intelligent mining of TCM oncology classics. Zhao et al[26] optimized the construction and data processing flow of the modern famous old doctor medical case database by using the HanLP word segmentation system and a hidden Markov model, significantly improving the quality and efficiency of data mining. Modern databases, including the Traditional Chinese Medicine Integrative Database, the Traditional Chinese Medicine Systems Pharmacology Database and Analysis Platform (TCMSP), the Encyclopedia of Traditional Chinese Medicine, SymMap, and HERB, as well as disease and molecular databases such as Online Mendelian Inheritance in Man, DrugBank, Therapeutic Target Database, Kyoto Encyclopedia of Genes and Genomes, and Gene Expression Omnibus, have been widely used in the construction of KGs for multi-dimensional correlations of “syndrome - symptom - prescription - herb - target” in malignant tumors such as breast cancer, liver cancer, gastric cancer, and colorectal cancer[27-29]. With the deepening of tumor immunotherapy research, the TCMIO database further integrated molecular mechanism data related to the immunological regulation of TCM, providing a specialized data platform for interpreting the scientific connotation of TCM regulating the tumor microenvironment[30].
Model architecture: Multi-algorithm adaptation for intelligent analysis of TCM data
Based on the standardized integration of multi-dimensional data, various DL and NLP models have become core tools for uncovering the hidden diagnostic and therapeutic rules of TCM in oncology. Distinct algorithmic models are adapted to specific research scenarios—such as ancient text review, medical case analysis, and association mining—forming exclusive architectures tailored to the unstructured data characteristics of TCM.
To address the unstructured and complex nature of clinical medical cases, existing research adopts the BERT-BiLSTM-CRF composite model to achieve structured decomposition and information extraction from lung cancer clinical records. By combining knowledge representation learning, this approach builds a link prediction architecture that accurately mines the hidden correlations between symptoms, tongue and pulse manifestations, TCM prescriptions, and tumor syndrome differentiation, thereby digitizing and intelligently extracting the diagnostic experience of renowned TCM practitioners[31]. Concurrently, the HanLP word segmentation system integrated with Hidden Markov Models efficiently processes large volumes of TCM medical case text data. This optimizes the workflow for data cleaning and knowledge mining, effectively resolving the challenges of low efficiency, strong subjectivity, and significant errors inherent in traditional manual medical case organization[26]. For ancient text data mining, an automated literature analysis framework accommodates the obscure semantics and unique expression patterns of classical Chinese, enabling batch retrieval, classified sorting, and visualized contextualization of ancient medical literature to construct a lightweight intelligent analysis model[25]. Ultimately, these multiple models are hierarchically adapted to the data characteristics of macro-theory, clinical diagnosis, and microscopic mechanisms, forming a comprehensive intelligent mining model system for TCM oncology.
Verification strategy: Multi-dimensional cross-evidence verification ensures the reliability of the research
To mitigate the risk of one-sided intelligent mining results or their deviation from clinical reality and TCM theories, existing studies suggest a multi-layer cross-validation strategy integrating ancient book tracing, clinical verification, and molecular evidence. This approach aims to enhance the scientific rigor and practical utility of the analytical results. First, based on foundational TCM classics such as the Yellow Emperor’s Inner Classic and the Treatise on Febrile Diseases, the framework ensures that all data mining and model analysis results can be traced back to the TCM tumor syndrome differentiation and treatment system, thereby aligning with the core theoretical connotations of TCM[25]. Second, relying on decades of real-world clinical medical records accumulated by renowned TCM masters (e.g., Zhong-Ying Zhou), these authentic clinical data were utilized for both model training and pattern mining. This ensures that the intelligent mining outcomes genuinely reflect actual clinical diagnosis and treatment scenarios[31].
Finally, through microscopic validation via tumor molecular biology and immune regulation mechanisms, the framework leverages target and pathway data from multi-omics databases. This verifies the molecular mechanisms by which TCM prescriptions, syndrome differentiation, and classification regulate tumor initiation and progression, as well as reshape the tumor microenvironment. Consequently, this establishes a robust three-layer cross-validation system: “Ancient book theory – clinical practice – molecular mechanism”[27-30]. By effectively preventing data bias and theoretical disconnect inherent in purely algorithmic mining, this strategy provides reliable technical assurances and systemic support for the precise intelligent research of tumors from an integrative Chinese-Western medicine perspective.
Syndrome differentiation and efficacy prediction: Clinical decision support integrating disease and syndrome
AI-driven syndrome differentiation and efficacy prediction models are enabling the intelligent upgrading of diagnostic and therapeutic decision-making. In syndrome differentiation, AI algorithms can perform in-depth mining and pattern recognition of information derived from the four diagnostic methods of TCM. Huang et al[32] used neural network models to standardize terminology and then combined them with cluster analysis to classify syndrome-related information extracted from EHRs of patients with breast cancer. Their results showed that liver-gallbladder dampness-heat syndrome was the predominant syndrome pattern and that the model outputs were largely consistent with physicians’ syndrome differentiation results. Ding et al[33] quantitatively processed symptoms, signs, and information from four diagnostic methods from EHRs of patients with primary hepatocellular carcinoma based on fuzzy mathematics and integrated extreme learning machine networks with particle swarm optimization algorithms to construct syndrome classification and prediction models. These models effectively captured nonlinear associations between clinical data and different syndromes, thereby providing new technical pathways for the standardization, objectification, and intelligentization of TCM syndrome differentiation. In efficacy prediction and risk stratification, Zhang et al[34] proposed a collaborative prevention and treatment framework for “pan-cancer extreme early stage” disease, integrating TCM and Western medicine. Centered on biological networks, this framework integrates clinical phenotypes, TCM syndromes, and multi-omics data, and uses AI-based dynamic modeling to achieve accurate prediction of tumor risk and optimization of combined TCM and Western medical treatment strategies. Tang et al[35] applied logistic regression, random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM) algorithms to develop prognostic models integrating TCM interventions and TCM syndrome characteristics for the early identification of recurrence and metastasis risk. When validated in 558 patients with stage I-III colorectal cancer undergoing radical surgery, the XGBoost model showed favorable discrimination for 3-year and 5-year recurrence and metastasis risk [area under the curve (AUC) > 0.75]. Fu et al[36] used retrospective cohort data to construct efficacy prediction models for non-small cell lung cancer (NSCLC). Multivariable regression analysis indicated that, after adjustment for confounding factors including chemotherapy, targeted therapy, and radiotherapy, TCM treatment remained significantly associated with a 26% reduction in mortality risk. By integrating the dynamic evolution of TCM syndromes with modern survival analysis and leveraging ensemble learning algorithms, it may be possible to develop individualized risk stratification tools with both strong predictive performance and clinical interpretability.
Active component screening and mechanistic elucidation of Chinese herbs: AI empowering new drug discovery
The integration of ML, network pharmacology, and multi-omics enables the systematic mining of key active constituents in Chinese herbs and the in-depth elucidation of their molecular mechanisms in regulating the tumor microenvironment. Liang et al[37] investigated Sangye, Heye, and Shaji and integrated TCMSP, SwissTargetPrediction, and other databases to screen key active components such as quercetin, kaempferol, and isorhamnetin. They then constructed disease-drug intersecting target networks in combination with weighted gene co-expression network analysis and differential expression analysis. Using SVM-recursive feature elimination, RF, and least absolute shrinkage and selection operator algorithms for screening and validation, they ultimately identified core targets for oral squamous cell carcinoma and esophageal squamous cell carcinoma, among which matrix metalloproteinase 1 was identified as a common cross-cancer hub gene closely associated with tumor invasion and metastasis. Guo et al[38] classified active components of licorice using unsupervised clustering algorithms and combined this approach with density functional theory calculations and molecular docking to identify key residues, including ASN364 and TYR485, in nitric oxide synthase 2. They further elucidated the structure-activity relationship of licorice chalcones against hepatocellular carcinoma and identified core pathways such as nuclear factor kappa B and programmed death ligand 1 (PD-L1)/programmed cell death protein 1 (PD-1), thereby opening new avenues for small-molecule drug design in liver cancer. Building on these advances, promoting the coordinated development of Chinese medicinal standardization systems, innovative drug research and development, and toxicity risk management has become an important direction for translating mechanistic studies of antitumor Chinese herbs into clinical applications. Through component fingerprint profiling, quality marker (Q-marker) identification, and consistency evaluation, standardized systems can be established covering medicinal material sources, processing procedures, and formulation quality, thereby consolidating the material basis of Chinese medicinal products[39,40]; combined with AI-based screening, organoid models, and integrated pharmacokinetic, pharmacodynamic, and toxicological evaluation, the research and translation of active constituents and compound candidate drugs can be accelerated[41,42]; meanwhile, the integration of toxicogenomics, metabolomics, gut microbiota analysis, and real-world data can facilitate toxicity risk identification and the construction of efficacy-toxicity balance assessment systems as well as whole-course safety early warning systems for Chinese herbs. These efforts provide a methodological foundation for the modernization of Chinese herbs and the discovery of novel antitumor drugs[43,44].
Optimization of whole-course diagnosis and treatment management: Integrated practice of TCM and Western medicine under a multidisciplinary team framework
In whole-course cancer management, AI can be incorporated into the multidisciplinary team (MDT) framework to integrate clinical characteristics, imaging, pathology, laboratory indicators, treatment responses, patient-reported outcomes (PROs), and TCM syndrome information, thereby establishing risk warning and decision-support systems that cover the perioperative period, systemic treatment period, and rehabilitation/follow-up period (Figure 2).
Figure 2 Timeline-based multimodal data modality and artificial intelligence intervention points across the whole-course management of cancer integrating traditional Chinese medicine and Western medicine.
TCM: Traditional Chinese Medicine. Created by Figdraw (ID: IUTIO0097a; Supplementary material).
During the perioperative period, the synergistic use of AI and enhanced recovery after surgery (ERAS) principles enables prediction of postoperative infection, delayed gastrointestinal recovery, pulmonary complications, and length of hospital stay by integrating age, comorbidities, nutritional status, inflammatory indicators, and surgery-related parameters, thereby facilitating individualized optimization of perioperative management[45]. At the same time, TCM has unique advantages in regulating stress responses, promoting gastrointestinal function recovery, alleviating pain, and improving overall functional status[46]. AI can further identify patient subgroups most likely to benefit from interventions such as Chinese medicinal therapy, acupuncture, and acupoint stimulation, thereby improving the implementation of ERAS pathways. Based on a meta-analysis of randomized controlled trials, Zhang et al[47] found that acupuncture significantly shortened the time to recovery of defecation, flatus, and bowel sounds in patients with postoperative ileus after colorectal cancer surgery. They further used the Apriori algorithm, frequent pattern growth algorithm, complex network analysis, and cluster analysis to identify Zusanli (ST36) and Shangjuxu (ST37) as high-frequency core acupoints, thereby providing a theoretical basis for optimizing acupuncture prescriptions during the perioperative period.
During chemotherapy, radiotherapy, targeted therapy, and immunotherapy, toxicity monitoring and supportive care are critical to maintaining treatment efficacy and adherence. AI can use EHRs, laboratory data, medication records, wearable devices, and patient-reported symptoms to enable the early identification and dynamic stratification of myelosuppression, nausea and vomiting, diarrhea, hepatic and renal dysfunction, peripheral neurotoxicity, and immune-related adverse events, thereby shifting management from passive response to proactive intervention[48]. Combined with real-world data on integrated TCM and Western medicine, AI can also identify associations between syndrome evolution and toxicity spectra, assisting in the optimization of Chinese medicinal prescriptions, acupuncture regimens, and nutritional support strategies to improve tolerance to systemic therapy. In a prospective cohort study, Sun et al[49] were the first to apply TCM tongue diagnosis to the prediction of chemotherapy-induced myelosuppression (CIM) risk in patients with lung cancer. They identified “swollen tongue” as an independent risk factor for CIM (odds ratio = 3.67) and combined it with the Karnofsky performance status score and the number of highly toxic chemotherapeutic agents to construct a nomogram with good discrimination (AUC = 0.82).
During the rehabilitation and follow-up stage, AI demonstrates additional advantages in continuous management. By integrating radiomics, circulating tumor DNA (ctDNA), laboratory indicators, lifestyle data, and long-term symptom monitoring, AI can dynamically assess the risk of recurrence and metastasis, the level of functional recovery, and changes in quality of life, and can accordingly optimize follow-up frequency and intervention pathways[50]. TCM at this stage emphasizes holistic recuperation, symptom control, and constitutional recovery, whereas AI can incorporate syndrome pattern changes, sleep, fatigue, mood, nutrition, and activity status into a unified model to support individualized TCM intervention and long-term health management[51]. Overall, AI-driven whole-course management based on integrated TCM and Western medicine is not merely a combination of two medical systems; rather, within the MDT framework, it enables synergy between disease diagnosis and syndrome differentiation, balances risk prediction with functional regulation, and integrates disease control with rehabilitation support, thereby providing cancer patients with more precise, continuous, and high-quality integrated care pathways.
Clinical translation of AI-empowered precision oncology
In cancer screening and early diagnosis, the clinical translation of precision medicine is primarily reflected in the identification of high-risk populations, the development of molecular biomarkers, and the optimization of detection strategies. Multidimensional data analysis methods based on radiomics and DL have been widely applied to pulmonary nodule detection, mammographic screening, and colorectal endoscopic lesion recognition, demonstrating considerable value in improving lesion detection rates, reducing the risk of missed diagnosis, and enhancing diagnostic consistency[52-54]. However, imaging-based assessment mainly relies on morphological characteristics and is insufficient to fully capture the molecular features and biological heterogeneity of tumors. The development of liquid biopsy has partially addressed this limitation and has become an important complementary strategy for early cancer screening. For genetically susceptible populations, germline testing for pathogenic variants in mismatch repair genes such as MLH1 and MSH2 has become well established for risk assessment in pancreatic cancer and Lynch syndrome-associated colorectal, ovarian, and endometrial cancers[55,56]. For sporadic tumors, integrated analytical models based on ctDNA methylation profiles, fragmentomic features, mutational spectra, and multimodal molecular biomarkers have demonstrated favorable sensitivity and specificity in the early detection of hepatocellular carcinoma, lung cancer, and colorectal cancer[57,58]. Meanwhile, advances in highly sensitive sequencing technologies have further improved the detection of low-abundance mutations, providing technical support for the reliable identification of ctDNA in early-stage tumors[59]. Together, these advances are facilitating a shift in cancer care from traditional symptom-driven, diagnosis-oriented approaches toward models centered on risk identification and early intervention[60].
In risk stratification and treatment decision support, precision medicine has progressively incorporated patient clinical phenotypes, tumor molecular subtypes, immune microenvironment characteristics, and relevant biomarkers to provide quantitative evidence for individualized therapeutic strategies. Clinical phenotype information, including age, tumor stage, histological type, prior treatment history, and therapeutic response, forms the basic framework for baseline risk assessment and initial treatment selection[1]. However, reliance on conventional clinicopathological parameters alone is insufficient to fully reflect tumor biological heterogeneity. Multi-omics integration can substantially improve the accuracy of tumor molecular subtyping and enable more refined classification of tumor subtypes. Yang et al[61] achieved 98.0% accuracy in predicting triple-negative breast cancer subtypes by integrating messenger RNA, microRNA, gene mutation, DNA methylation, and magnetic resonance imaging radiomics data, markedly outperforming single-omics models. Beyond molecular characteristics, the composition, spatial distribution, and histomorphological features of immune cells within the tumor immune microenvironment also provide critical information for risk assessment and therapeutic optimization. With digital pathology and DL, these features can be extracted in a high-throughput manner from routine hematoxylin and eosin-stained sections. In gastrointestinal malignancies, DL has enabled the direct prediction of microsatellite instability (MSI) status from pathological images, offering a cost-effective and efficient approach for identifying populations likely to benefit from immunotherapy[62]. The HER2 quantitative scoring model for gastric cancer established by Han et al[63] further demonstrated the potential of digital pathology in biomarker identification and decision support. On this basis, the integration of key biomarkers, including epidermal growth factor receptor, anaplastic lymphoma kinase, HER2, and MSI, enables more accurate selection of beneficiary populations and optimization of therapeutic strategies, thereby promoting the evolution of cancer care toward individualized management[64].
In treatment response evaluation and prognostic management, the advantages of precision medicine are mainly reflected in real-time molecular monitoring during therapy, early identification of resistance mechanisms, dynamic tracking of minimal residual disease (MRD), and the construction of long-term prognostic models. Compared with traditional efficacy evaluation methods, which primarily depend on imaging-based changes in lesion size and improvements in clinical symptoms, AI algorithms can integrate serial ctDNA data and clinical variables to build dynamic risk scoring systems that more precisely capture temporal heterogeneity in treatment response. For MRD monitoring, Chaudhuri et al[65] used cancer personalized profiling by deep sequencing for dynamic ctDNA monitoring in 40 postoperative patients with stage I-III NSCLC, demonstrating that molecular residual disease was detectable before radiographic confirmation in approximately 72% of relapsed patients, with a median lead time of 5.2 months. This finding provides a critical window for early clinical intervention and adjustment of therapeutic strategies. For long-term prognostic assessment, Yuan et al[66] used NLP-assisted machine learning to extract clinical information from EHRs and construct an overall survival (OS) prediction model for NSCLC, achieving AUC values of 0.828, 0.825, 0.814, 0.814, and 0.812 for predicting 1-year, 2-year, 3-year, 4-year, and 5-year OS, respectively, significantly outperforming traditional clinical assessment methods. The multimodal DL framework UMPSNet, developed by Zhang et al[67], further overcame the limitation of traditional prognostic models restricted to a single cancer type and achieved the transition from single-cancer prognosis prediction to pan-cancer survival prediction. By integrating PROs, wearable device monitoring data, and long-term follow-up records, ML-based prognostic models are gradually expanding from the prediction of single survival outcomes to the comprehensive assessment of recurrence patterns, metastatic trajectories, and treatment-related toxicities, thereby promoting the evolution of cancer care from static evaluation to dynamic, continuous, and whole-course management.
Barriers in the integrated digital-intelligent system: Inherent defects and practical constraints of AI-enabled TCM oncology research
Currently, the application of AI in TCM is facing two core challenges: Deep heterogeneity and incommensurability. At the data level, the information from the four diagnostic methods of TCM, tongue and pulse signs, as well as the data on the combination of prescriptions and drugs, exhibits significant spatiotemporal heterogeneity and strong non-linear characteristics. There is a severe shortage of high-quality, standardized, labeled datasets. The existing disease-specific databases mostly remain at the shallow digitalization stage of electronic medical records, lacking structured and computable governance at the semantic level, and thus are unable to support the in-depth mining and model training of AI algorithms. At the algorithm level, the thinking of syndrome differentiation and treatment, and the compatibility logic of monarch, minister, assistant, and envoy herbs are highly dependent on the implicit cognition of physicians. They cannot be formalized into learnable and verifiable deterministic rules, resulting in insufficient generalization ability and clinical stability of the models. At the trust level, the inherent “black box” attribute of DL hinders mutual verification with TCM theories, and AI models lacking causal reasoning capabilities are difficult to gain acceptance and recognition from clinical physicians. Constrained by these bottlenecks, current related research is still limited to single-center, retrospective methodological validation. Building an intelligent decision-making system for TCM that can be embedded in the clinical diagnosis and treatment loop requires crossing the dual gaps of theoretical adaptation and engineering implementation to achieve forward-looking and large-scale implementation capabilities[68].
CONCLUSION
Realistic challenges and prospective outlook for integrated traditional Chinese medicine and Western medicine in precision oncology.
Despite the substantial technical support provided by AI for integrated TCM and Western medicine in precision oncology, this field remains at a critical stage of transition from conceptual exploration to systematic development. From theoretical integration to data governance, and from methodological standardization to the generation of clinical evidence, collaborative precision diagnosis and treatment involving TCM and Western medicine still face multilevel challenges. Further breakthroughs are urgently needed in conceptual frameworks, technical pathways, and translational models.
First, data heterogeneity and insufficient standardization remain major obstacles to deep integration. TCM diagnostic and therapeutic information, including tongue manifestations, pulse patterns, and syndrome differentiation, is highly subjective, dynamically variable, and often unstructured. In contrast, Western oncology data are typically represented as high-dimensional quantitative data, such as imaging, pathology, and multi-omics data. The substantial differences between these two types of data in terms of acquisition methods, representation formats, and underlying cognitive paradigms make it difficult for AI models to establish unified, stable, and clinically meaningful representations[69]. Second, existing AI models still rely heavily on “black-box” algorithms such as DL, whose internal decision-making processes are difficult to explain either through the logic of TCM syndrome differentiation or in terms of modern biomedical mechanisms. In high-risk scenarios such as cancer diagnosis and treatment, insufficient interpretability directly affects clinical trust, application safety, and the feasibility of large-scale implementation[20]. Third, the current evidence base remains insufficient to support the standardized application of integrated TCM and Western medicine in precision oncology. Most existing studies are single-center, retrospective, and based on small sample sizes, with endpoints mainly focused on symptom relief, short-term toxicity improvement, or local therapeutic response. High-quality studies addressing key outcomes such as OS, progression-free survival, recurrence, and metastasis risk, and long-term quality of life are still lacking. In particular, TCM interventions are highly individualized and dynamically adjusted, making conventional randomized controlled trials (RCTs) less well-suited to their complex intervention patterns and thus limiting the strength and external generalizability of the evidence[70]. Fourth, multimodal fusion and dynamic modeling capabilities still require further improvement. Tumor evolution is essentially the result of continuous interactions between tumor biological processes and host responses, whereas TCM syndrome differentiation and treatment emphasize dynamic adjustment according to timing, location, and individual constitution. Therefore, data from a single time point or a single dimension are insufficient to fully capture the complete disease spectrum. Most current models remain at the stage of static integration or shallow feature concatenation and cannot effectively characterize disease-state transitions, treatment-phase changes, or long-term evolutionary patterns[71]. Finally, ethical, regulatory, and data security issues will continue to influence the sustainable development of this field. Precision oncology involves highly sensitive data, including genetic information, treatment records, adverse events, and long-term prognostic outcomes. In the context of integrated TCM and Western medicine, the addition of large volumes of unstructured text and individualized diagnostic and therapeutic information further increases the requirements for data de-identification, sharing mechanisms, and cross-institutional data flow[72].
In the post-pandemic era, biomedical research has entered a new stage characterized by data ecosystem reconstruction and the deep integration of multiple disciplines. The data-intensive scientific paradigm has gradually come to dominate the direction of the field’s development. As integrated governance systems for the entire data lifecycle and intelligent technologies continue to merge, they are constantly reshaping the research logic and expanding the application boundaries of biomedicine. This provides a highly favorable technical environment and a solid research foundation for the innovative development of precision oncology that integrates traditional Chinese and Western medicine[73]. Furthermore, with the continuous acceleration of interdisciplinary convergence, the synergistic enabling effects of AI and modern biomedical science have increasingly come to the fore. The cross-integrative model utilizing multimodal data from both TCM and Western medicine is becoming progressively mature. This approach is poised to continuously innovate traditional research systems in scenarios such as tumor screening, precise prevention, and individualized diagnosis and treatment, thereby promoting the evolution of integrated TCM and Western medicine diagnostic models toward greater intelligence and refinement[74].
Looking ahead, based on the full implementation of Findable, Accessible, Interoperable, Reusable data principles and Health Level Seven Fast Healthcare Interoperability Resources standards, it will be possible to further break down the barriers posed by multi-source heterogeneous medical data. This will facilitate data standardization, sharing, and reusability; effectively reduce the costs associated with data integration and preprocessing; and continuously improve the efficiency of transforming medical data into actionable knowledge. Ultimately, these advancements will lay a solid data foundation for the large-scale research and clinical application of integrated TCM and Western medicine in precise tumor diagnosis and treatment[75]. Simultaneously, under the stringent requirements of medical data security compliance and cross-border, cross-center scientific collaboration, by leveraging cutting-edge technologies such as distributed control, trusted computing, and federated learning and deeply integrating them into the medical governance system, researchers can establish a robust scientific collaboration framework. This framework, characterized by the principle of “keeping data localized while enabling collaborative model iteration”, will provide secure and stable infrastructural support for method innovation and model iteration in integrated TCM and Western medicine tumor research[34].
Overall, the integrated traditional Chinese and Western medicine approach to cancer precision medicine is still in the exploratory and integrative stage. Moving forward, efforts can be further intensified in four core directions: Standardization, dynamic modeling, mechanism elucidation, and clinical evidence translation.
At the data system level, it is imperative to refine the terminology norms, data standards, and knowledge representation systems specific to integrated TCM and Western medicine. Addressing the industry pain points of poor compatibility of heterogeneous data and insufficient computability will pave the way for the efficient integration and standardized application of diverse medical data. At the model construction level, dynamic predictive models covering the entire tumor course can be developed based on multimodal clinical data. These models will accurately identify patient populations that benefit most from specific treatments, providing a quantitative basis for individualized and differentiated clinical interventions. At the mechanism research level, by integrating multi-omics sequencing, single-cell analysis, and network pharmacology, researchers can systematically elucidate the core targets and molecular mechanisms of TCM compound therapies. This approach will fill the existing gaps in the ambiguous and insufficient theoretical support governing the integration of TCM and Western medicine. At the clinical translation level, promoting the complementary integration of RCTs and real-world studies will help construct an evidence-based system that balances scientific rigor with clinical applicability. Furthermore, deeply embedding interpretable intelligent assistive decision-making tools into EHRs and multidisciplinary diagnosis and treatment processes will effectively bridge the chasm between basic research and clinical application.
With improved data standards, continuous innovation in research methods, and a more mature evidence-based system, the integrated TCM and Western medicine approach to cancer precision medicine will ultimately transcend its current fragmented and exploratory state. It is expected to gradually establish a novel, standardized clinical diagnosis and treatment paradigm that is verifiable, implementable, and scalable. This may provide crucial support for innovation in cancer precision diagnosis and treatment, as well as for the modernized development of integrative medicine.
