From Data-Driven Cities to Data-Driven Tumors: Dynamic Digital Twins for Adaptive Oncology

Irem Karaman^1*Burhan Sebin²

• ¹University of Miami Health System, Miami, United States
• ²Vice President, Atlas Space, Miami, United States

IntroductionOncology is undergoing a transformation due to the advent of digital twin technology, which enables precision therapy by creating synchronized virtual copies of physical systems. Unlike static models, dynamic digital twins continually integrate multimodal patient data such as clinical, imaging, and molecular data, to simulate therapy scenarios and direct real-time therapeutic decisions (1). The success of smart city digital twins, which control complicated systems like traffic and energy in real time, provides a model for oncology (2). Like cities, tumors are dynamic, multiscale systems shaped by therapy, immunological responses, and genetic alterations. There must be less trial-and-error in cancer care if we can use dynamic, learning-based twins to practice therapy instead of static ones that rely on pre-treatment snapshots, which miss this progression.The Limits of Static Oncology TwinsStatic oncology twins are often created from a single pre-treatment dataset, which includes imaging and genetic profiles, and used to predict first medication responses. However, when tumors adapt into their environment and therapy, these one-time models lose accuracy, failing to predict emergent resistance mutations and microenvironmental modifications (3). Static twins are unable to detect early signals of relapse or toxicity because they exclude longitudinal indicators such as ctDNA kinetics, biomarker signatures, and routine labs, as well as real-time physiological data from wearables (4). Moreover, while assays like Oncotype DX and spatial cell‐type mapping inform initial risk stratification, they remain disconnected from iterative clinical decision loops (5). Integrative-cluster trials categorize patients based on a combination of molecular and histopathologic characteristics; however, they still do not incorporate closed-loop adaptation. On the other hand, a dynamic twin would actively adapt therapy by recalibrating its AI and mechanical models with each new data stream. Nonetheless, prospective validation via observational cohorts and randomized studies is necessary to establish the clinical utility of these adaptive systems. Smart Cities as Operational TemplatesDigital twins in smart cities serve as central command centers for urban ecosystems, consistently integrating data from traffic cameras, smart streetlights, water distribution monitors, public transit GPS, and air quality sensors (6). They function as a comprehensive citywide system, executing real-time simulations to evaluate the impacts of diverse scenarios, including the closure of a highway during peak hours, the reconfiguration of traffic signals to mitigate congestion, or the redirection of emergency response vehicles in unusual weather conditions. The results are then utilized to provide optimized control signals to the physical infrastructure (7). For example, Virtual Singapore uses real-time environmental and transportation data from over 30 agencies to simulate urban planning, emergency response, and energy efficiency, making it a global benchmark for centralized scenario rehearsal (8). The open digital twin of Helsinki facilitates public participation and climate planning via accessible 3D simulations, aiding in the visualization of solar potential and guiding zoning choices (9). Simultaneously, Shanghai's urban management twin enhances everyday operations and emergency response by combining IoT, AI, and real-time monitoring across districts, resulting in up to 30% gains in municipal efficiency (10). Upon the activation of a flood alert, the twin can promptly simulate reservoir releases, road closures, and diversion routes, analogous to how an ICU dashboard predicts patient stability across various ventilator settings (Figure 1).An oncology twin reflects this dynamic, feedback-oriented urban model. The system assimilates ongoing clinical notes and imaging data, analogous to how a city twin processes CCTV feeds (11,12). It incorporates pathology reports, genomic sequencing, and biomarker trends, similar to the way environmental sensors monitor particulate levels. Heart rate, blood pressure, sleep patterns, stress levels, changes in glucose levels, and even gait analysis are all wearable streams that can be used like mobile noise or air quality tools to let the twin know about changes in the body (13). In order to allow predictive algorithms to make real-time adjustments to dosing regimens or therapy switches, the city's infrastructure is routinely audited using serial liquid biopsies and tissue samples. By considering the tumor and its host as an interdependent and dynamic system, similar to simulating rush-hour traffic and power-grid load, the twin may predict spikes in tumor growth, mutations that confer resistance, and areas of toxicity that are particularly harmful. This provides doctors with a real-time practice ground for therapy and individualized treatment plans.Case Studies: Dynamic Twins in ActionRecent efforts demonstrate the feasibility of dynamic oncology twins, with new use cases expanding their scope:Table 1. Representative dynamic twin case studies in oncology and smart city examples highlighting continuous data ingestion, scenario simulation, and adaptive feedback.Digital‑Twin ApplicationDomain / SettingLead Institution(s)Core Data & ModelReported Outcome / InsightReferenceGlioblastoma Radiotherapy TwinNeuro‑oncologyOden Institute, UT AustinBayesian tumour‑growth model updated with serial MRI during radiotherapyAdaptive dosing delayed median progression by ≈6 days with lower total dose (14)Lung‑Cancer 3‑D TwinThoracic oncologyStanford + NCI–DOEDeep CNNs on CT, digital pathology, genomics; ctDNA updatesReconstructs 3‑D tumour, infers EGFR status, forecasts resistance, suggests therapy switches (15)FarrSight Virtual‑Trial TwinsMulticancer trial optimisationConcr (VISION trial)Patient‑specific simulation of alternative regimensHigher response when real therapy matched twin recommendation (16)Melanoma Immunotherapy TwinImmuno‑oncologyIndiana UniversityMultiscale agent‑based immune–tumour modelPredicts immune escape; tests checkpoint sequencing (17)Pain‑Management TwinSupportive care / PK‑PDMulti‑center Population PK/PD model of fentanylOptimizes dosing, reduces adverse events (18)Personalized AML Chemo SchedulerHematologyAcademic consortiumLongitudinal counts + mechanistic kineticsAvoided leukopenia in 10/13 AML cases (19)Breath‑Gas Early‑Detection TwinNon‑invasive screeningIndustry–academic groupVolatile metabolite ML signaturesDetects pre‑clinical tumor shifts via breath (20)Clinical‑Trial Design TwinTrial optimizationMulti‑center Virtual cohorts predicting toxicity vs efficacyReduced adverse events in trial simulations (21)Smart‑City Traffic Twin Urban operationsCity of Singapore – 'Virtual Singapore'Live traffic, IoT, weather feeds + agent‑based modelReal‑time rerouting cut congestion by ~15 % (8)Shanghai Municipal Digital TwinUrban operationsShanghai Municipal Government; Smart City ProgramCity wide IoT sensor network, AI analytics, real time cross district monitoringUp to 30 % gain in municipal efficiency and faster emergency response (10)These case studies highlight dynamic twins’ core functionalities: real-time data assimilation, multiscale modeling, and therapy rehearsal, extending to education and trial design.Digital twins in oncology must integrate patient data from diverse sources alongside the expertise of oncologists, clinical guidelines, and pertinent decision-making criteria. This ensures that the twin can demonstrate decision-making processes in a more intricate manner, particularly when it must evaluate individual critical criteria rather than solely relying on raw data streams. Incorporating domain-specific knowledge, such as the relative importance of prognostic signs or patient comorbidities, into digital twin recommendations ensures that they are consistent with established therapeutic rationale. Instead of adopting a one-size-fits-all approach, oncology digital twins should be created for each type of cancer since tumors evolve and treatments vary. For example, a glioblastoma twin must consider the tumor's dissemination and its responsiveness to radiation therapy, whereas a breast cancer twin would emphasize the functionality of hormone receptors and the malignancy's sensitivity to chemotherapy (14). Conversely, lung cancer twins consider mutational profiles such as EGFR or ALK status when determining targeted therapy options (15). Numerous patient data exist within EHR/EMR systems; nevertheless, the generation of real-time digital twins is challenging due to data silos, inconsistent formatting, absent longitudinal records, and delays in data acquisition. Confronting these difficulties requires the synchronization of data streams, the establishment of interoperability standards, and the assurance of real-time data that is readily available. Contemporary clinical monitoring systems also have challenges in accurately modeling tumor development, medication resistance, and treatment-related toxicity at the individual level. Challenges involve the limited resolution of conventional imaging techniques, insufficient liquid biopsy collection, and inadequate real-time biomarker surveillance, including circulating tumor DNA or PD-L1 fluctuations (21). More precise and dynamic depictions of disease progression and treatment effect are made possible by new technology that are rapidly improving, such as high-frequency wearable biosensors, serial liquid biopsy platforms, and advanced imaging methods.Dynamic digital twins in cancer care, which evolve in response to evolving patient data, are distinct from static ones that only run once. FarrSight®-Twin perpetually integrates novel genetic variants, does repeated whole-slide scans, and incorporates time-stamped clinical events, thereby recalibrating its predictive model with each update topredict the responses of breast cancer patients to treatment and immunotherapy (16). The Stanford–NCI-DOE lung cancer twin integrates follow-up CT images, interval pathology samples, and novel genetic and clinical data through a systematic process. Each update recalibrates the tumor growth and treatment response trajectories, transforming the model from a static representation into a dynamic virtual patient (15). During each MRI session—T1-contrast, T2-FLAIR, and diffusion—the Bayesian engine assimilates the new voxel-level contours and ADC measurements, updates the coefficients for each patient's proliferation, invasion, and radiosensitivity, and subsequently recalculates iso-dose maps and fractionation. This loop transforms the UT Austin glioblastoma twin into a dynamic therapeutic guide, capable of adjusting treatment intensity in response to emerging infiltrative areas or reducing dosage upon confirmation of tumor shrinkage (14). This enables physicians to formulate therapy protocols that are more efficacious in decelerating the disease and mitigating its damage.Technical Foundations of Dynamic Oncology Twins The foundation of every interactive digital twin is a solid data flow. For oncology, this entails combining several patient data sets into a single repository, such as EHRs, lab findings, radiographs (CT, MRI, PET), histopathology reports, liquid biopsies (circulating tumor DNA), and multi-omics datasets (genomics, transcriptomics, proteomics). Complying with established standards is essential for achieving interoperability. These standards include the OMOP Common Data Model for observational health data and the HL7 Fast Healthcare Interchange (FHIR) for clinical and imaging metadata (22). Additionally, automated extraction workflows and streaming APIs guarantee that new patient measurements are transferred into the counterpart with minimal latency, thereby maintaining real-time fidelity to the changing disease state (23). Upon establishment of these streams, Apache Kafka facilitates the real-time transfer of clinical notes, imaging files, and multi-omics results, encapsulated in FHIR, OMOP, DICOMweb, or Phenopackets formats (24). The incoming messages are stored in an RDF database that associates each data point with terms from SNOMED CT, LOINC, OncoKB, and NCIt, followed by the application of a variational auto-encoder to address any gaps (25). Subsequently, modality-specific AI models operate with high efficiency: 3D UNet++ and DenseNet-121 for CT/MRI, a Swin-Transformer for whole-slide images, LoRA-tuned DNABERT-2 for genomic variants, a Temporal Fusion Transformer for irregular lab series, Hetero-GraphSAGE for knowledge graphs, physics-informed networks for tumor growth equations, and a PPO agent that weighs projected survival benefits against toxicity (26). Hybrid models that integrate physics-informed and data-driven approaches utilize multimodal embeddings, including cross-attention early fusion, Bayesian late fusion, and tensor-gated hybrid fusion. They disseminate calibrated uncertainty, enabling approximately 1,000 therapy-rehearsal simulations to conclude in under 1 second, accompanied by median projections and 95% prediction intervals (27).A flexible modeling system that integrates data-driven AI with mechanistic simulations is equally essential. Convolutional neural networks and transformer models can derive predictive characteristics from imaging and genomic sequences, respectively, whereas systems of ordinary and partial differential equations represent tumor development dynamics and drug–tumor interactions (28). Agent-based models mimic microenvironment dynamics and immune-cell infiltration; physics-informed neural networks apply biological limits on acquired representations (29). Generative methods, like variational autoencoders and generative adversarial networks, let you do "what-if" studies by putting together virtual groups of people who would be treated differently.The scenario simulation engines that are built on top of these models serve as platforms for the rehearsal of virtual therapy. The twin predicts important results including tumor shrinkage, resistance emergence, and toxicity profiles by listing potential treatment plans, which may include different medication combinations, dosage regimes, or sequence orders (16). Utilizing reinforcement-learning algorithms allows for the optimization of therapeutic methods in pursuit of multi-objective goals, such as maximizing progression-free survival while minimizing side effects. New patient data is constantly being used by these algorithms to update policy judgments. Lastly, simulation at the point of care must be scalable and have minimal latency, and this can only be achieved with a solid computing foundation.