Deep neural networks are an innovative technology, and applications within medical research are continually emerging. We understand that regulators and clinical trial sponsors may be less familiar with these technologies than other statistical techniques. 

Despite this, we've experienced positive regulatory interactions, particularly for PROCOVA^TM, our core method behind TwinRCTs, which has been qualified by the European Medicines Agency (EMA) for primary analysis in phase 2 and 3 trials. In addition, the US Food and Drug Administration (FDA) has stated that it concurred with the EMA's qualification opinion. These regulatory interactions, including positive FDA feedback about our methods heard in collaboration with our sponsors, affirm PROCOVA as an acceptable methodology in compliance with current EMA and FDA guidelines.

In our engagement with regulatory bodies, we emphasize (i) transparency around model construction and training, (ii) thorough performance evaluations of models on independent test datasets, and (iii) pre-specification of model parameters and code for generating participants’ digital twins for trial analyses. Additionally, at the request of regulators, we offer tools like Shapley values to aid in model interpretability.

Quality control is paramount in preparing the training datasets, with meticulous attention to managing outliers, duplicates, and missing data. This ensures that digital twins generated for clinical trials are robust and reliable, even in the face of missing inputs during a trial. Variables for DTG inputs are selected based on their prevalence in related clinical trials, though adjustments are made to exclude data that do not meet quality standards.

When standard datasets do not fully align with the target patient group or a rare disease with limited data, Unlearn can finetune or customize DTGs. This involves adjusting the training data or applying methods like synthetic data and transfer learning to develop effective models despite data scarcity. Collaborations and data-sharing initiatives are key to acquiring enough relevant data for these rare conditions.

The effectiveness of digital twins is using diverse datasets ensuring that trial populations are well-represented in our training data. Regular updates with new training data and machine learning techniques are crucial, especially when the standard of care changes, ensuring that digital twins remain accurate and relevant.

Digital twins are generated using digital twin generators (DTGs), which are mostly trained on individual participant data from past clinical trials and observational studies. These data consist of detailed, longitudinal clinical records of nearly 1 million participants across more than 30 indications, including demographic information, clinical outcomes, labs, and biomarkers, tailored to each specific indication. This broad dataset ensures a comprehensive coverage of a wide spectrum of patient demographics, disease stages, and treatment histories sourced globally. Such extensive data allow DTGs to closely align with the characteristics of target patient groups and specific trial requirements, enhancing their ability to generalize to new situations and patient populations.

Digital twin generators are advanced AI models that generate precise predictions of patient-specific clinical outcomes. Unlike traditional covariate adjustment methods that rely on broader, less individualized data points, participants’ digital twins achieve unparalleled correlation with actual observed outcomes. By integrating these detailed, probabilistic forecasts into our analysis, we are able to significantly enhance the statistical power of clinical trials. We believe this is only the beginning of AI’s potential to transform clinical research.

TwinRCTs use prognostic scores from participants' digital twins as 'super covariates' to substantially decrease variance in treatment effect estimates. This innovative approach allows TwinRCTs to either maintain their statistical power with fewer participants in the control group compared to traditional trials or to achieve higher power without enlarging the sample size. By integrating digital twins and employing the PROCOVA methodology, TwinRCTs are designed to ensure that potential biases in the forecasts made by digital twins still maintain unbiased estimation of treatment effects.

Unlike TwinRCTs, clinical trials with external controls use data from non-enrolled patients to estimate treatment effects, similar to synthetic controls, which use statistical methods to select external patient data for comparison. While external and synthetic controls offer alternatives in scenarios where randomized controls are impractical, they frequently encounter regulatory scrutiny due to concerns over data comparability and potential biases.