Artificial intelligence in laboratory medicine: From machine learning to large language models.
TL;DR
It is concluded that AI is best positioned, in the near term, as an assistive layer rather than an autonomous agent, and the rationale for developing vertically oriented, laboratory-specific foundation models that can integrate structured results, longitudinal patient data, quality-control metadata, instrument information, and clinical knowledge is discussed.
Abstract
ABSTRACT The integration of artificial intelligence (AI) into laboratory medicine has undergone a remarkable evolution over the past decade, shifting from traditional machine learning (ML) applied to structured laboratory data to the recent emergence of large language models (LLMs) capable of processing unstructured clinical text, interpreting complex laboratory findings, and supporting clinical decision-making. This narrative review traces that progression and critically compares four main AI paradigms accessible to laboratory professionals: conventional ML, deep learning, general-purpose LLMs, and laboratory-specific foundation models. We examine representative application domains-blood cell morphological analysis, autoverification, acute infectious risk stratification, interpretation of urinalysis data, clinical decision support, and report generation-and assess the maturity of the evidence for each domain. Laboratories contemplating AI adoption face a distinctive set of challenges: preanalytical variability, interplatform calibration differences, effects of the specimen quality, reagent-lot sensitivity, and the context dependence of result interpretation, all of which demand validation practices beyond those that are standard in general medical AI research. We conclude that AI is best positioned, in the near term, as an assistive layer rather than an autonomous agent. Its deployment requires a clear problem definition, disciplined data engineering, external validation, prospective silent trials, human-in-the-loop oversight, and postdeployment surveillance aligned with existing laboratory quality systems. We additionally discussed the rationale for developing vertically oriented, laboratory-specific foundation models that can integrate structured results, longitudinal patient data, quality-control metadata, instrument information, and clinical knowledge. Clinical accountability must remain with qualified laboratory professionals.