AI Creates Living Digital Twins to Personalize Cancer Radiotherapy

A groundbreaking artificial intelligence system is transforming how radiation therapy is delivered to lung cancer patients by creating dynamic digital twins that evolve alongside each patient's unique biological response. Published on arXiv on February 15, 2026, the research introduces COMPASS (Comprehensive Personalized Assessment System), which represents a paradigm shift from population-based treatment models to truly personalized adaptive radiotherapy.

The Limitations of Current Radiotherapy Practice

Modern radiotherapy for non-small cell lung cancer (NSCLC) has become increasingly precise, generating vast amounts of data from each treatment session. Current biologically guided radiotherapy (BGRT) regimens capture metabolic, anatomical, and dose information with each fraction delivered. However, clinical decision-making remains largely informed by static, population-based normal tissue complication probability (NTCP) models that overlook the dynamic, patient-specific biological trajectories encoded in sequential treatment data.

"Traditional models average out individual biological responses," explains the research team. "They fail to capture the unique temporal evolution of each patient's tissues during treatment, potentially missing early warning signs of toxicity."

How COMPASS Creates Living Digital Twins

The COMPASS system functions as a temporal digital twin architecture that integrates per-fraction PET scans, CT imaging, dosiomics, radiomics, and cumulative biologically equivalent dose (BED) kinetics. Unlike static models, COMPASS treats normal tissue biology as a dynamic time series process, creating a living digital representation that updates with each treatment session.

At the core of the system is a gated recurrent unit (GRU) autoencoder that learns organ-specific latent trajectories from the sequential data. These trajectories capture how organs like the spinal cord, heart, and esophagus respond to radiation over time. The system then employs logistic regression to classify these trajectories and predict eventual CTCAE grade 1 or higher toxicity.

Early Results from Initial Patient Cohort

The study analyzed eight NSCLC patients undergoing BGRT, contributing to 99 organ fraction observations covering 24 organ trajectories. Despite the small cohort size, the intensive temporal phenotyping allowed for comprehensive analysis of individual dose response dynamics.

The findings revealed a viable AI-driven early warning window, with increasing risk ratings occurring several fractions before clinical toxicity became apparent. This advance could give clinicians crucial days to adjust treatment plans before patients experience adverse effects.

Perhaps most significantly, the dense BED-driven representation revealed biologically relevant spatial dose texture characteristics that occur before toxicity and are typically averaged out with traditional volume-based dosimetry. These subtle patterns in how radiation is distributed within tissues appear to be critical predictors of biological response.

Implications for Adaptive Radiotherapy

COMPASS establishes a proof of concept for AI-enabled adaptive radiotherapy, where treatment is guided by a continually updated digital twin that tracks each patient's evolving biological response. This represents a fundamental shift from reactive to proactive treatment management.

"The system moves us from population averages to individual trajectories," notes the research. "Each patient's digital twin becomes more accurate with each treatment fraction, potentially allowing for real-time treatment adjustments that maximize tumor control while minimizing side effects."

Technical Implementation and Validation

The GRU autoencoder architecture was particularly suited to this application due to its ability to handle sequential data and capture temporal dependencies. By learning compressed representations (latent trajectories) of the high-dimensional treatment data, the system could identify patterns that would be invisible to human observers or traditional statistical methods.

The logistic regression classifier achieved promising results despite the limited dataset, suggesting that with larger patient cohorts, the predictive accuracy could improve substantially. The researchers emphasize that while the current study focused on NSCLC, the approach is theoretically applicable to other cancer types treated with radiotherapy.

Future Directions and Clinical Integration

The COMPASS system opens several important avenues for future research and clinical implementation. Larger validation studies are needed to confirm the early warning capabilities across diverse patient populations. Integration with existing radiotherapy planning systems will be crucial for clinical adoption.

Additionally, the digital twin approach could be expanded to incorporate other data streams, such as genomic information, circulating biomarkers, or patient-reported outcomes, creating even more comprehensive models of individual treatment response.

Ethical and Practical Considerations

As with any AI system in healthcare, COMPASS raises important questions about validation, transparency, and clinical responsibility. The black-box nature of some deep learning components may present challenges for regulatory approval and clinician trust. The researchers acknowledge these concerns and suggest that future work should focus on making the system's predictions more interpretable to clinicians.

From a practical standpoint, implementing such systems would require significant infrastructure investments in data storage, processing capabilities, and clinician training. However, the potential benefits in terms of reduced toxicity and improved treatment outcomes could justify these investments.

Conclusion: Toward Truly Personalized Cancer Care

The COMPASS system represents a significant step toward realizing the promise of personalized medicine in oncology. By creating living digital twins that evolve with each patient's unique biological response to treatment, it moves beyond the one-size-fits-all approach that has dominated radiotherapy for decades.

As radiotherapy continues to become more precise and data-dense, AI systems like COMPASS will be essential for translating this wealth of information into improved patient outcomes. The early warning capability demonstrated in this study could fundamentally change how radiation toxicity is managed, potentially sparing countless patients from unnecessary suffering while maintaining or even improving tumor control.

The research, while preliminary, points toward a future where each cancer patient's treatment is guided by their own constantly updating digital twin—a future where medicine is not just personalized in theory, but dynamically adapted to each individual's evolving biological reality.