Share of US hospitals with predictive AI in electronic health records rises to 71%
In the United States, 71% of non-federal acute care hospitals reported using predictive artificial intelligence integrated with electronic health records in 2024. A year earlier, this figure was 66%, MyJoyOnline reports, citing data presented in the article.
The article notes that the reliability of such systems depends not only on algorithms but also on the quality, meaning, security and traceability of medical data. A completed medical record can be clinically misleading: a diagnosis used for billing does not always reflect the full picture of a patient's condition, a laboratory result may contain an incorrect unit of measurement, and a timestamp may record when information was entered into the system rather than the medical event itself.
Data exchange does not eliminate errors
The author draws attention to the problem of information-system interoperability. The FHIR standard makes it easier to represent and exchange medical data, but it cannot correct errors in source information or reconcile different definitions. Two systems may successfully transmit the same value but interpret it differently.
More current news is available on the UA.News Telegram channel Telegram.
The US Centers for Medicare & Medicaid Services rule on interoperability and prior authorization requires payers covered by it to implement or update a number of FHIR application programming interfaces. Compliance deadlines for these interfaces generally begin in 2027.
Models require continuous monitoring
AI systems must be tested in the specific conditions in which they will be used. Studies of models for detecting pneumonia in chest X-rays showed that models trained on data from one hospital system often performed worse on data from other healthcare facilities. Results may be affected by differences in patient populations, disease prevalence, equipment, workflows and coding practices.
In 2024, 79% of hospitals using predictive AI reported some form of post-implementation evaluation. However, not all of them assessed most or all of their models, and 18% of respondents did not know whether such monitoring had been carried out. Evaluation should cover not only accuracy but also calibration, predictive value, performance for different patient groups, external and temporal validation, and the impact on clinical processes.
