Abstract:
|
Artificial intelligence (AI) is increasingly being incorporated into diagnostic medical devices for intended uses such as screening, diagnosis, prognosis, risk prediction, clinical decision support, and phenotype classification. However, machine learning (ML)-based diagnostic devices present challenges to regulatory authorities evaluating their safety and effectiveness. First, due to inadequate training, an ML model may lack generalizability across patient subsets because of confounding or datapoints being overly influential in model fitting. Second, AI/ML has the potential to improve performance by learning from real-world use and experience, yet the regulatory pathway to fulfill this potential in a least burdensome manner needs further development. Other challenges include model transparency and interpretability, reuse of validation data, errors in the reference standard used to determine ground truth, risk calibration, and uncertainty quantification of device predictions. After reviewing FDA initiatives supporting the development and validation of ML based diagnostic devices for clinical use, we will review some of the statistical challenges and propose some solutions.
|