All Times EDT
Keywords: Machine Learning, device, companion diagnostic, SamD
The use of Machine Learning (ML) and advanced analytical approaches are becoming increasingly common in addressing key clinical questions and providing valuable insights for better patient management, understanding of treatment outcomes and potentially being able to select patients who have higher likelihood of responding to experimental therapies. The application of these methods can lead to development of companion diagnostic tools that gets reviewed through the FDA’s software as a medical device (SaMD) clinical evaluation process. While, early on in a development program, it is difficult to anticipate which exploratory hypotheses will ultimately be supported by data, there is value in taking certain steps to ensure rigor is maintained during the initial evidence generation process of such hypotheses. Development and validation of any diagnostic tool should be carefully planned to ensure no loss in development time and level of evidence required by regulatory agencies is sufficient to get the device approved. This can include adequate pre-specification of scientific hypothesis, careful selection of training and validation datasets during model development, and up-front discussion with regulators on amount of prospective evidence needed. The latter may require designing well powered prospective studies to evaluate benefit of device or the drug-device combination. Additionally, there needs to be consideration of the black-box element of ML models and how transparency can be maintained in communication with physicians and patients on how the tool works. We will discuss these issues using a case study in immuno-oncology.