PS32: Recent Advances in Digital Health Device Development and Validation: Statistical Strategies for Expanding Intended Uses

Digital health devices-including diagnostic imaging systems, image-guided interventional tools, and software-based radiological aids-are rapidly evolving as advances in digital technologies, artificial intelligence (AI), and data integration transform how imaging data are generated, analyzed, and used in clinical practice. These innovations enable novel device intended uses that challenge traditional statistical paradigms for development, validation, and regulatory decision-making. Some of the newest digital health devices are designed to automatically rule out diseases or conditions in negative cases without human review, placing a premium on controlling false-negative rates and on establishing reliable performance in low-prevalence settings. Risk prediction devices analyze extensive image databases to generate patient risk information, where validation is complicated by resource limitations and the inability to feasibly establish ground truth for all outcomes. In addition, lesion localization devices aim to identify the spatial location of abnormalities and provide a level of suspicion, requiring careful statistical treatment of multi-reader variability, spatial accuracy, and clinically meaningful performance metrics. More recently, large language model (LLM)–enabled radiological devices automatically generate suggestions for physicians by integrating medical images with patients' historical reports, introducing new challenges related to explainability, uncertainty quantification, and evaluation of text-based outputs derived from complex, multi-modal inputs. Across these intended uses, digital health devices increasingly rely on high-dimensional outputs, adaptive algorithms, and heterogeneous data sources. Validation is further complicated by imperfect or subjective reference standards, evolving software components, increasing number of functions integrated in the devices, and differences in imaging protocols, patient populations, and clinical workflows. Ensuring robust, transparent, and reproducible evidence in this setting requires close collaboration among statisticians, clinicians, engineers, data scientists, and regulators. This session will highlight recent advances in digital health device development with an emphasis on statistical methodologies tailored to diverse device intended uses, including rule-out, risk prediction, lesion localization, and LLM-enabled decision support devices. Speakers from regulatory agencies, industry, and academia will discuss practical study designs, performance metrics, and validation strategies that support impactful transdisciplinary decision-making in a rapidly changing digital landscape.