From prediction to reality: Five years of AI in healthcare. Adoption, impact, and the road ahead.
Authors
Affiliations (3)
Affiliations (3)
- Cardiff School of Technologies, Cardiff Metropolitan University, Llandaff Campus, Western Avenue, Cardiff CF5 2YB, United Kingdom; Balsamee Ltd, 1 Capital Quarter, Tyndall Street, Cardiff CF10 4BZ, United Kingdom. Electronic address: [email protected].
- Cardiff School of Technologies, Cardiff Metropolitan University, Llandaff Campus, Western Avenue, Cardiff CF5 2YB, United Kingdom.
- Balsamee Ltd, 1 Capital Quarter, Tyndall Street, Cardiff CF10 4BZ, United Kingdom.
Abstract
The 2020 EIT Health & McKinsey report Transforming Healthcare with AI was among the most cited forecasts shaping expectations for AI adoption in healthcare. It predicted early gains in administrative automation and medical imaging, then remote monitoring and natural language processing (NLP), and eventually integrated clinical decision support. To assess how accurately those predictions tracked real-world developments through the end of 2025, and to identify where they held, fell short, or were overtaken. This structured narrative review compares the report's domain-level forecasts against 2020-2025 evidence from peer-reviewed implementation studies, U.S. FDA regulatory data, national and international guidance, industry surveys, and foundation-model evaluations. Each prediction was classified as realised, under-realised, exceeded, or unanticipated against explicit criteria. Two predictions held: administrative automation and medical imaging led early adoption, with ambient documentation scaling from pilots to deployment and more than 1300 AI/ML-enabled devices authorised by the FDA, mostly in radiology. Two under-delivered: remote monitoring and classical NLP decision support, constrained by interoperability, reimbursement, and workflow barriers. Procurement accelerated faster than forecast, though chiefly on industry rather than peer-reviewed evidence. The largest divergence was the emergence of generative AI and multimodal foundation models from late 2022, which, with the COVID-driven acceleration of wearables, fell outside what the forecast could reasonably have anticipated. The roadmap identified the right domains but misjudged the pace of discontinuous change and overestimated the readiness of data, governance, and workflow infrastructure. Progress toward 2030 will depend less on new model architectures than on interoperability, governance designed for generative systems, implementation science, workforce readiness, and business-model design. Emerging agent-protocol standards warrant empirical evaluation as routes to addressing long-standing interoperability gaps. Without those foundations, scaled and equitable deployment will remain out of reach.