Owkin receives EMA letter of support for innovative AI approach to oncology trial analysis

Owkin has received a letter of support from the European Medicines Agency (EMA) for its innovative approach to building prognostic covariates by applying deep learning to histology slides for oncology trials.

What is Covariate Adjustment and why are they important in clinical trials?

Clinical trials are crucial to developing new cancer treatments, but they are expensive and time-consuming. When trials fail, it is often hard to know whether it was due to drug ineffectiveness or suboptimal trial design. Covariate adjustment can increase the probability of success of trials.

Covariate adjustment is a technique used during the statistical analysis of a trial to adjust for variables that are prognostic i.e. that are correlated to the clinical endpoint of the study. By removing the variation of the endpoint explained by the covariates, researchers can more accurately measure the treatment effect of the intervention being studied. This helps to increase the statistical power of the study and can improve the accuracy and reliability of the results. However, the current approach to selecting these covariates is unsystematic, with medical experts choosing them based on their prior knowledge or reusing covariates from similar trials. This method can reduce the trial's statistical power if important covariates are omitted. To improve this, researchers require a novel approach to selecting prognostic covariates in oncology trials.

Watch our covariate adjustment explainer video here:

‍

What is different about Owkin’s approach?

Owkin’s data-driven approach leverages access to research-grade multimodal data to identify prognostic covariates for use as adjustment covariates in each specific indication. This approach is at the core of our multi-year strategic partnership with BMS.

While a more systematic approach to the selection of covariates can already deliver strong value to clinical trial sponsors, AI can unlock further prognostic information in imaging modalities. In the case of today’s announcement, we have leveraged digital pathology to better predict survival in two cancer types and we are excited that the EMA has written a letter of support for the use of those covariates as adjustment covariates in clinical trials.

What is the topic of the letter?

The letter describes the use of two of our deep learning models, MesoNet (published in Nature Medicine) and HCCnet (published in Hepatology), for predicting overall survival in mesothelioma and resected Hepatocellular carcinoma (HCC) patients. The models were trained using digitized pathology images and a large pre-trained model to extract information from images, focusing on small regions of the biopsy that are most relevant to prediction. The models were trained on data from the MESOPATH/MESOBANK database and Henri Mondor Hospital and validated using testing datasets from The Cancer Genome Atlas (TCGA).

How do we propose to use the predictive models?

We are proposing to use the predictions of these two models as prognostic biomarkers for the adjustment of efficacy analysis on overall survival of life-prolonging drugs in randomized phase II and phase III clinical trials. Specifically, Mesonet predictions are proposed to be used for life-prolonging drugs in first-line malignant pleural mesothelioma patients, while HCCnet predictions are proposed for drugs in the adjuvant setting for resected hepatocellular carcinoma patients.

‍

‍

What are the benefits and drawbacks of these predictive models?

The main benefit of those models is:

The gain in prognostic performance compared to an adjustment with covariates used in current practice in clinical trial settings. This gain in performance could translate to gains in statistical power.

The EMA also acknowledges other strengths of the proposed models such as the robustness to artefacts implied by the attention mechanism, the interpretability of the models, and the removal of inter-observer variability through automation. However, the EMA highlighted that the evidence base for the AI model approach as new technology is limited compared to traditional approaches such as histological subtyping and may be sensitive to technical biases.

‍

‍

What is the impact of adjustment on those deep-learning covariates?

Owkin has evaluated the impact of covariate adjustment in the time-to-event setting in a simulation study. In particular, we show that adjustment on Mesonet or HCCnet on top of traditional covariates reduces the sample size requirements by more than 10% for trials in the corresponding indications. The EMA noted that those simulations are under the proportional hazards assumption which may limit their generalisability.

What is the EMA position?

The EMA recognizes the appropriateness of the approach: “A strong pathophysiological rationale for including imaging information as prognostic information is plausible”. But as this is a new technology, more validation is required to reach a Qualification Opinion.  Consequently, the letter encourages us to “use our approach in future trials and to do additional prospective validation”. This Letter Of Support will help us convince pharmaceutical companies to partner with us on those next steps.

‍

‍

Félix Balazard, Director of Optimized Development at Owkin comments:

This EMA letter of support offers a clear regulatory path for the integration of AI models based on histopathology into the analysis of randomized clinical trials. Owkin is bringing AI to the regulatory setting.

What are Owkin’s next steps?

The use of traditional covariate adjustment in clinical trials has been endorsed by regulatory agencies like the EMA and FDA. This letter represents an important step towards integrating deep learning models into trial design, which can significantly improve the success rates of trials. Ongoing Mesothelioma and HCC trials can benefit from this approach, and for other indications, we can develop on-demand prognostic scores using our Federated Research Network. Our expertise in developing accurate prognostic models and our access to high-quality data and medical experts make us a valuable partner in reducing clinical trial failure rates. We welcome partnerships to validate and improve our approach. Adopting data-driven covariate adjustment can accelerate progress towards new treatments and ultimately lead to better patient outcomes.

In addition to the two AI models presented in this article, which have been designed for use in clinical trials, we are dedicated to developing regulatory-grade AI models that are based on digital pathology. Our commitment to this area is demonstrated by our successful development of two AI diagnostic solutions, as well as our recent involvement in the PortrAIt consortium. This consortium aims to develop and deploy 15 AI-based tools in clinical settings. By continuing to advance AI technology, we hope to improve patient outcomes and make a meaningful impact in the field of healthcare.