Artificial intelligence (AI) is transforming health care by accelerating advances in medical product development and tools for augmenting the capabilities of health care practitioners. In particular, the number of FDA-approved, AI-enabled medical devices has exploded from a handful in 2010 to nearly 1250 today, with nearly 1000 in radiology and over 100 for cardiovascular use (https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices). Approved AI-enabled medical devices include those intended for imaging acquisition and processing (e.g., segmentation, denoising, reconstruction), classification and diagnosis, early detection of life-threatening, time sensitive conditions (e.g., intracranial hemorrhage or sepsis), and long-term risk assessment (e.g., 5-year breast cancer risk). AI-enabled devices on the horizon include multi-modal or other large language model (LLM) software as a medical device (SaMD) for generating clinical reports intended for medical decision support, individualized decision support via digital twins, and robotics to assist with endoscopy or other critical and time sensitive medical procedures. Evaluation of the development and performance of AI-enabled medical devices can appear daunting for many reasons. For example, AI models underlying devices may be developed on a vast amount of clinical data and / or synthetic data, have a huge effective number of parameters, be unexplainable and uninterpretable, and be designed to generate complex output for an unfamiliar intended use. In this talk I will review recurring statistical issues with AI-enabled device development and performance evaluation. For these issues, I will attempt to simplify matters by applying general statistical principles with the hope of identifying either an existing statistical method for addressing the issue or a needed statistical innovation. Opportunities for statistical innovation in medical device AI, if seized, may lead to more predictable, transparent, and efficient practices and tools for development and evaluation, which in turn may shorten the expected review times of AI-enabled devices by regulators, thereby allowing patients and providers more timely access to these products.