Applying ML to medical devices and radiology imaging

Unlike traditional software, ML models evolve over time and improve as they interact with real-world data. Also, due to the probabilistic nature of the models, it is likely that the output of these models will change as the statistics behind the data shift. This poses a challenge in applying these models for regulated medical workflows because the medical decision-making process needs to be consistent and supported by the same evidence over and over again. Moreover, the results of an ML model aiding in a clinical decision-making process need to be explainable. In other words, we cannot treat the model as a “black box”; we need to understand its inner workings and explain its behavior in specific scenarios.

In spite of these challenges, the FDA recognizes that AI/ML has the potential to transform healthcare due to its ability to derive insights from vast amounts of data generated in healthcare practice every day....