Cancers are complex adaptive dynamic systems, and the underlying mechanisms of such systems are best understood when subjected to perturbation. Therapies such as radiation, chemotherapy, or immunotherapy, among others, introduce a substantial perturbation to a tumour, which offers the opportunity to measure response, identify putative vulnerabilities, and dynamically adapt treatment as necessary. A framework that could facilitate adaptive therapies is the so-called digital twins. Digital twins aim to mimic the structure, context, and behaviour of the physical twin—the cancer patient in oncology. Digital twins are dynamically updated with data from the patient, have predictive capability, and inform clinical decision-making.1 This enables the comparison of different treatment options for individual patients through in silico trials beyond the average outcomes of historic clinical trials (Figure 1).2 The power to make predictions for individual patients comes from mathematical models that are validated with the data of past patients, and iteratively calibrated to the individual with the patient's own data.3 While digital twins are established in different industries such as mechanical and aerospace engineering,4 their use in medicine remains to be demonstrated. One major limiting factor is the lack of continuous, accurate measurements of all variables of interest—such as tumour growth dynamics, drug concentrations, complete blood counts, patient-reported outcomes, among others. Thus, digital twins in oncology will unlikely mimic the whole patient, but rather simulate limited subsystems of the patient that realise value for clinical decision making—such as the tumour, or the vascular system, or organs at risk. For each of these subsystems of interest, a predictive mathematical model (or an ensemble of multiple predictive mechanistic mathematical models) needs to be developed, calibrated, and validated, and its predictive power needs to be assessed.5 Once sufficient prediction accuracy is demonstrated on historic data (how accurate such models need to be may depend on the clinical problem, data uncertainty, and physician need), prospective clinical studies need to evaluate the clinical utility—that is, can the digital twin framework (1) iteratively use newly collected clinical data to retrain its predictive models and (2) simulate factual and (3) counterfactual treatments, and (3) communicate model predictions with (4) associated uncertainties and (5) model accuracy (6) in real time to (7) inform the physician and patient interactions with the likelihood of success of different treatment strategies. To evaluate the clinical utility of digital twin and mathematical model-guided personalized adaptive therapy, the responses and outcomes of such framework-guided therapy needs to be evaluated against either the counterfactual standard of care response and outcome predictions for an individual patient, or the outcomes of patients in a digital twin-treated trial arm compared against the outcomes of patients treated with physician's choice therapy without model-guided adaptations. To integrate and adopt mathematical models and digital twins into clinical decision-making, suitable use cases need to be identified. There is a plethora of mathematical models that propose to “optimise” therapy. Optimal therapy, however, is not a common clinical need, and the entry barriers for untested clinical treatment doses and protocols are incredibly high. Furthermore, optimal therapy may look very different for different patients, and the optimisation objectives are likely to change for each individual patient during the course of treatment. Of urgent clinical need is the identification and selection of the treatment that has the highest likelihood of success from the growing number of clinically approved therapies for the given cancer type, stage, and treatment history.6 Model-guided treatment selection within standard of care, or for clinical decisions of advanced cancers without any established standard of care protocols, may provide opportunities to demonstrate the utility of predictive modelling in oncology. One example of the former may be the prospective validation of the proliferation saturation index (PSI) model7 for radiation fractionation selection in head and neck cancer. Different radiation fractionation protocols are considered standard of care for head and neck cancer, yet to date, no biomarker exists to select which protocol would yield the most robust tumour volume reduction for individual patients. The PSI model predicted that patients with slow-growing tumours (high PSI) would be better treated with hyperfractionation (that is, radiation twice a day with a dose of 1.2 Gy per fraction), whereas patients with faster-growing tumours (low PSI) would be controlled with once daily radiation with a dose of 2 Gy per fraction. The validation trial of 57 patients demonstrated a 12% improvement in patients achieving a robust mid-treatment response—thereby providing the first prospective evidence that mathematical modelling may be used to guide personalised radiation fractionation.8 Feasibility of predictive modelling for patients without standard of care has been successfully demonstrated in 15 patients as part of the Evolutionary Tumour Board at Moffitt Cancer Center, a non-interventional prospective pilot study in which novel therapeutic strategies based on evolutionary principles were developed for individual patients.9 Using an extension of the tumor-growth-inhibition (TGI) model,10 evolutionary therapy-based treatment strategies could be developed and presented to the treating physician in clinically realistic time in 11/15 patients (73%). The physician and patient followed the model recommendation in all 11 cases, demonstrating that the model and evolutionary tumour board workflow can aid in clinical decision making.9 The above samples of prospective clinical trial validation of predictive models share two important commonalities: (1) they were driven by specific clinical questions and commissioned by clinicians, and (2) both the PSI and the TGI model are mathematically simple. As these models had to be calibrated on limited patient-specific data, biological and mathematical complexity had to be sacrificed in favour of calibratability. Not despite their simplicity, but because of their simplicity, were these models clinically actionable. While, without a doubt, more complex models would simulate more biological variables and dynamics and be mathematically more interesting, the above-mentioned models were fit for purpose—the prerequisite for integrating and adopting mathematical models and digital twins into clinical decision making. This work was supported in part by NIH/NCI U01CA244100 and U01CA280849. The author is inventor on Patent No.: US 12,406,772 B2, Systems and methods for predicting individual patient response to radiotherapy using a dynamic carrying capacity model. Not applicable.
Heiko Enderling (Fri,) studied this question.