Pragmatic considerations for fostering reproducible research in artificial intelligence

Abstract

Artificial intelligence and deep learning methods hold great promise in the medical sciences in areas such as enhanced tumor identification from radiographic images, and natural language processing to extract complex information from electronic health records. Scientific review of AI algorithms has involved reproducibility, in which investigators share protocols, raw data, and programming codes. Within the realm of medicine, reproducibility introduces important challenges, including risk to patient privacy, challenges in reproducing results, and questions regarding ownership and financial value of large medical datasets. Scientific review, however, mandates some form of resolution of these inherent conflicts. We propose several approaches to permit scientific review while maintaining patient privacy and data confidentiality.

Artificial Intelligence in Cardiology: Present and Future

“…106,107 Although traditional research studies are positioned to share readily deidentifiable data, AI is now using large volumes of data that are much harder, potentially impossible, to deidentify. 108,109 Outside the context of the regulatory and ethical aspects of sharing health care data, there is also a question of the health care data's value and ownership. Health care organizations may have invested substantial sums of money in acquiring the technical personnel and equipment required to annotate data sets for AI research.…”

Quantitative Molecular Positron Emission Tomography Imaging Using Advanced Deep Learning Techniques

“…The AI and ML/DL community has witnessed groundbreaking algorithmic developments, and many of the algorithms are available to end users and developers as open-source software that can be further evaluated and validated for potential deployment in clinical and research settings. However, as reported by Hutson (125), only a small fraction of AI research papers provide code or pseudocode, which is prejudicial to reproducible research (126). Leading professional societies [e.g., MICCAI Society, Radiological Society of North America, American Association of Physicists in Medicine, Society of Nuclear Medicine and Molecular Imaging (SNMMI)] often organize international challenges for validation and comparison of competitive medical image analysis algorithms, leading to the creation of rankings and guidelines regarding the performance of these approaches under controlled conditions by using public image repositories (127).…”

Ethical Use of Electronic Health Record Data and Artificial Intelligence: Recommendations of the Primary Care Informatics Working Group of the International Medical Informatics Association

“…While recommended, scientific review of AI algorithms for reproducibility through sharing of protocols, raw data, and programming codes, introduces risk to patient privacy. It also raises questions about ownership and financial value of large real world datasets and systems [33]. The transformation to AI-enhanced health care needs to be judicious, informed, and systematic to ensure safety and quality of data and care during the transition.…”

“…This sharing can be achieved through the use of videos showing screenshots with outputs based on specific queries or allowing access to "data enclaves" where the tester can drive the AI-based system. Whether this balanced approach is achievable will require a degree of trust within the scientific community that appropriate development and evaluation methods were used [33].…”