Prospective real-world implementation of deep learning systems in healthcare: a systematic review guided by implementation science.

Abstract

Deep learning (DL) applications in healthcare are expanding beyond proof-of-concept studies. Yet, the extent of its real-world implementation and impact on patient care and clinical workflows remains unclear due to the limited prospective real-world findings. Understanding how DL tools perform in real clinical environments is critical for guiding successful and sustainable deployment. Using a layered methodology with established implementation science frameworks, this systematic review aimed to systematically map the implementation strategies and outcomes of prospective DL implementation studies, proposing recommendations based on identified gaps of relevant studies to serve as a guide for the future implementation of DL systems. 20 articles were included: 3 from radiology, 1 from otolaryngology, 3 from dermatology, and 13 from ophthalmology. All studies assessed clinical outcomes, demonstrating the effectiveness and feasibility of integrating DL systems into existing clinical workflows. Adoption and appropriateness were the most frequently evaluated implementation outcomes; only one study evaluated implementation costs, and none evaluated sustainability. Stakeholder acceptability was only evaluated in 8 studies. Given the paucity of real-world DL implementation research, continued research into the clinical deployment of DL systems using hybrid effectiveness-implementation study designs as a framework is essential to facilitate its seamless and effective adoption into clinical practice.