Digital Twins for Radiation Oncology

Patient data is the central component of a digital twin. The data needed to establish an accurate digital twin will at least include patient demographic and summarizing data, which can be used to fill in any other data gaps with demographic-specific averages. Digital twins will also require three-dimensional tomographic imaging (CT, MR, etc.) to provide spatial data about the patient's tissues and anatomy. There will also be a need to transition from this physical spatial data (such as x-ray attenuation in CT images, magnetic spin density in MR images, etc.) to biological spatial data (such as tissue density, oxygenation values, histological profiling, etc.). Recent advances in multiscale modeling are currently exploring this possibility, and multiscale modeling is likely to advance in the future [7]. Digital twins may also incorporate genomic data, which can have a strong impact on treatment outcomes. In the future, genomic data may be used to augment the capabilities of multiscale modeling to provide better spatial tissue stratification.

Digital twin predictions will be enabled by a fusion of biophysical modeling and machine learning. This modeling approach can incorporate analytically derived models for microbiological processes, such as proliferation, cell death, cell repair, angiogenesis, mutation, and immune system responses, with hyperparameters for these models determined by modern deep learning methods. The potential to use this combination inference strategy has recently been demonstrated [8, 9, 10, 11]. Once equipped with patient data, a digital twin can use these models to move forward in time to predict the results of treatment decisions. By directly simulating patient progress rather than predicting results indirectly, this strategy is likely to provide models with improved predictive performance. Importantly, digital twin simulations can enable Monte-Carlo-like probabilistic simulation, where many predictions are made to sample the probability distribution of possible treatment outcomes.

The computing power required to sustain a digital twin is likely to be very high. As explained above, digital twins for radiation oncology will require patient data with dense sampling across many patients, and the sheer volume of this data alone will be computationally intensive to process. Recent research has demonstrated the impact of this volume of data, as well as the need for high-performance computing to manage it [12]. This data will be even larger when accounting for the model-specific parameters that are generated, as well as the results created when propagating the patient data forward in time repeatedly during treatment optimization or virtual clinical trials (see sections 5.1 and 5.2 below). For the universal database (see section 3), this data is multiplied by the number of patients. Beyond the data processing requirements, model training and validation for both the multiscale models and machine learning involve intense algorithmic complexity. To address these requirements, high-performance computing will be needed. This computing power can be centralized along with the universal database; therefore, individual institutions will not be required to provide their own computing power.

The primary benefit of digital twins for cancer patients is the ability to make more precise predictions for the course of each patient. While a large amount of cancer research has focused on predicting the probabilities of patient outcomes and side effects using artificial intelligence or statistical models, much of this research attempts to predict the results directly from the patient's imaging and phenotyping. As a result, these models are "synthetic", i.e. they ignore all of the complex biological and physiological phenomena that fundamentally determine the patient's progress. Digital twins innovate the current scientific paradigm by reflecting these biological and physiological not just from AI or statistics but also through mechanistic modeling, genetic, environmental, and social factors. In doing so, digital twins progress away from synthetic inference and move closer to actual simulation of the patient. This makes digital twins much more likely to correctly predict the patient's outcome trajectory.

Digital twins can fundamentally change the paradigm of biological optimization for radiation therapy treatment planning. Historically, biological optimization relies solely on tumor volume and the delivered radiation dose distribution, discarding much of the information that influences cancer radiobiology such as oxygenation and vasculature density. With digital twins, biological optimization can properly consider these factors, possibly greatly improving the resulting treatment plans. Similarly, digital twins can provide much more power to adaptive radiation therapy, a treatment paradigm in which the patient's radiation course is updated over the radiation course to account for the patient's anatomical changes. By using digital twins, adaptive radiation therapy can not only improve radiation dose homogeneity and conformality due to changes, but it can also modify the course based on the changing probabilities of cancer cure and toxic side effects. This can profoundly improve patient outcomes from radiation therapy in a previously impossible way.

Beyond the immense clinical side effects, digital twins can also be powerful tools for cancer researchers. Because of their low-level simulations of patients, digital twins are much easier to generalize beyond their training data. This makes it easy for researchers to investigate numerous possible combinations and ratios of therapeutic agents and strategies. Even further, digital twins will be able to extend their predictions to novel therapeutic strategies by simulating the patient's reaction to them. This goes beyond the typical artificial intelligence limitation of being unable to predict results that are significantly different from the training data. Because of their virtual nature, digital twins could enable large-scale virtual clinical trials that are more thorough and drastically faster than current clinical trials, accelerating the translation from basic science to clinical solutions.

Historically, individual hospital systems have maintained their own IT departments and databases in the interest of preserving the security of their patient health information. Therefore, it is likely that the database powering the digital twins would be maintained by the individual hospital systems that are using the digital twins to improve the outcomes of their own patients. While this makes it easier for patient information to be secure, hospitals would be hampering the effectiveness of their digital twins by forcing their digital twins to learn from only the hospital's data. Artificial intelligence and machine learning models improve significantly when their training data increases, so we expect that digital twins should become much more accurate and powerful if they were trained from the data of many hospital systems rather than one. Moreover, this format would require each hospital system to have their own high-performance computers in order to train and maintain hospital-specific digitial twin models. This increases the difficulty and cost of implementing digital twins on the hospital system's end.

However, with modern cryptography techniques and cloud computing readily available, it would be possible for hospital systems to anonymously share their patients' de-identified data with some central organization while still preserving the electronic key to that patient. This would allow the hospital system to periodically update the universal database with each patient's progress and clinical decisions, continuously improving the universal database by enabling online learning techniques [13]. The universal database would also be able to house the computers required to train the models without requiring high-end computing power on the side of the hospital systems. This would make it much easier and less costly for each hospital system to start using digital twins for their own patients, which incentivizes more hospital systems to join and even further improve the universal database.

Regardless of how the digital twins are maintained and trained, they will require an assembly of patient-specific data to make patient-specific inferences. Some of this data may be commonly acquired in typical patients, such as blood panels or MRI or CT imaging of the relevant part of the body. Depending on the final technical implementation of the digital twins, less-common data may need to be included, such as genetic sequencing of relevant genes. Even with the assistance of a central universal database, novel data commons and assembly methods need to be developed to minimize the number of additional tests needed to generate a digital twin. Moreover, each digital twin would operate better with more of its patient's data, so steps should be taken to increase the density of pre-existing patient data. For example, a digital twin would be able to produce better predictions for a patient undergoing lung radiation therapy if that patient has previously undergone radiation therapy. In this example, it would be ideal for the hospital to access all the data associated with the prior radiation therapy, but this would be difficult if that data were secured in another hospital system's database. Naturally, this problem also vanishes in the presence of universal databases because the hospital would be able to provide the patient's anonymized key to the database and retrieve all the patient's historical data, maximizing the created digital twin's predictive performance.

Although artificial intelligence and digital twin technology have been applied broadly to other industries, they have not been extensively investigated in the low-level human domain generated by multiscale modeling. This is primarily because the relevant advancements in multiscale modeling are relatively new. Therefore, there has not been enough time or awareness of researchers to develop models in these domains yet. The domains also pose unique traits due to their representation of lower-level human data. These traits open the door to a new class of artificial intelligence models that exploit the well-known mechanical, chemical, and biological properties of materials. This will enable the path from high-level empirical models that ignore the fine details of the human body to fundamental, physically motivated models that perform realistic human simulation. These models are an active area of research, but because they are so new and underdeveloped, the coupling of modern AI techniques with mechanistic models poses a unique challenge to digital twin development. However, it is likely that significant progress will be made on these models in the near future.