Regulation of AI algorithms for clinical decision support: a personal opinion

• Published in: International journal for computer assisted radiology and surgery Vol. 19; no. 4; pp. 609 - 611
• Main Author: Kandarpa, Kris
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.04.2024; Springer Nature B.V
• Subjects: Algorithms; Anticoagulants; Artificial Intelligence; Certification; Clinical decision making; Computer Imaging; Computer Science; Data sources; Datasets; Decision Support Systems, Clinical; Deep learning; Editorial; False information; FDA approval; Geographical distribution; Health Informatics; Humans; Imaging; Labels; Medical device industry; Medical equipment; Medical technology; Medicine; Medicine & Public Health; Multiculturalism & pluralism; Neural networks; Pattern Recognition and Graphics; Radiology; Regulation; Regulatory agencies; Software; Surgery; Trust; Version control; Vision; United States > US
• ISSN: 1861-6429; 1861-6410; 1861-6429
• DOI: 10.1007/s11548-024-03088-0

Meticulous provenance of data [3], from whom, where [including imaging infrastructure and protocols], and when data were collected, along with documentation of population diversity [gender, racial, ethnic, socioeconomic, appropriate age distribution] within geographic areas should be non-negotiable. ‘Time-stamped’ datasets, being important for monitoring temporal drifts in the contained data [e.g., due to changes in disease prevalence, population migration, technology or standards of clinical care], would provide auditable trails to enable version control and replicability, while enhancing the generalizability of an algorithm. A concrete method developers may use during deployment is to provide data source ‘nutrition labels’ [like content labels on packaged food] specifying the geographic diversity, representation [gender, age, racial, ethnic, socioeconomic, etc.] and collection periods of data used for training and validation [5]. [...]regulators may require developers to reveal the characteristics and limitations of the data source and declare optimal ‘conditions’ under which the submitted CDSA should be employed.