Healthcare data unreadiness

Health systems are finding that most of the data inside their own walls is not ready to run clinical AI. Zach Evans, chief technology officer at Xsolis, wrote on April 20 that less than 20% of enterprise data is usable for AI without major preparation. (hitconsultant.net) The basic problem is simple: artificial intelligence needs data that is consistent, labeled, and comparable across systems. Evans said many hospitals still rely on older electronic health record and laboratory information system setups built for billing, documentation, and transaction processing rather than pattern-finding by machine-learning models. (hitconsultant.net) An electronic health record is the software clinicians use to document, store, retrieve, share, and analyze patient-care information. A laboratory information system manages lab requests, specimen tracking, results reporting, and related workflow, which means both systems generate large volumes of data that AI tools later try to reuse. (healthit.gov) (ncbi.nlm.nih.gov) The snag is the labeling around that data. Metadata is the descriptive information that tells a system what a field means, how it was created, what format it uses, and how it should be interpreted; without it, a model can read the value but miss the context. (csrc.nist.gov) (niaid.nih.gov) Evans gave concrete examples of what goes wrong. A temperature field can mix Fahrenheit and Celsius with no unit attached, diagnosis entries can appear as free text instead of standardized codes, and medication names can alternate between brand names, generic names, and internal hospital codes. (hitconsultant.net) Those mismatches show up fast in pilot projects. Evans wrote that teams sometimes spend longer explaining why a script produced the wrong answer than they would have spent doing the task by hand, and he said risk models can end up flagging nearly every patient as high risk when they cannot tell active conditions from resolved ones. (hitconsultant.net) Hospitals are pressing ahead with AI anyway. The Office of the National Coordinator for Health Information Technology said in September 2025 that 71% of hospitals reported using predictive AI integrated with the electronic health record in 2024, up from 66% in 2023. (healthit.gov) The data bottleneck is not limited to hospitals. Gartner said on February 26, 2025 that 63% of organizations either did not have or were unsure they had the right data-management practices for AI, and it predicted that 60% of AI projects unsupported by AI-ready data would be abandoned through 2026. (gartner.com) That is why the work Evans described looks less like model tuning and more like plumbing. He argued that standardizing fields, cleaning records, and governing metadata have to happen before health systems can make reliable claims about whether a clinical AI tool is accurate, safe, or worth scaling. (hitconsultant.net) The immediate question for health systems is not whether they can buy another AI product. It is whether they can make their own patient, lab, and operational data legible enough for the tools they already want to test. (hitconsultant.net)