Loading field map
Clinical informatics emerged from a persistent tension: how to systematically translate the vast and growing volume of clinical data and medical knowledge into tangible improvements in patient care. Since the 1960s, the subfield has developed nine major frameworks, each offering a different answer to that question. Some focused on building the infrastructure to capture and organize data; others on encoding expert knowledge into rules; still others on mining patterns from large datasets or on understanding the human and organizational contexts in which these systems operate. The history of clinical informatics is not a simple story of one framework replacing another, but a layered evolution in which earlier approaches often remain active, providing infrastructure or posing unresolved challenges that later frameworks must address.
The first framework, Clinical Information Systems and Electronic Health Records (1967–Present), established the foundational layer. Pioneering efforts at institutions like the Regenstrief Institute and the LDS Hospital created the first computer-based patient records, moving clinical documentation from paper to digital form. These early systems were not merely digital filing cabinets; they were designed to support clinical workflows and to make patient data accessible for both care and research. The central contribution of this framework was the creation of a structured, longitudinal record that could serve as a substrate for all subsequent analytical and decision-support activities.
Almost simultaneously, Rule-Based Clinical Decision Support (1967–Present) emerged as the first systematic attempt to embed medical knowledge directly into clinical workflows. The landmark MYCIN system at Stanford demonstrated that a computer could use a set of if-then rules derived from expert clinicians to recommend treatments for bacterial infections. This framework treated clinical knowledge as something that could be explicitly encoded and applied algorithmically. It coexisted with the early EHR systems, often as a separate module that would generate alerts or suggestions based on data entered into the record. The rule-based approach was transparent—clinicians could inspect the logic behind a recommendation—but it was brittle: maintaining and updating the rule sets required constant expert input, and the systems struggled to handle the complexity and variability of real-world clinical scenarios.
A third framework, Clinical Knowledge Representation and Terminology Engineering (1970–Present), addressed a deeper problem that the rule-based systems had exposed. If clinical data were to be used by computers, the meaning of terms like "myocardial infarction" or "type 2 diabetes" had to be precisely defined and linked across different systems. This framework developed controlled vocabularies (such as SNOMED CT) and knowledge representation models (such as the Unified Medical Language System, UMLS) that allowed clinical concepts to be encoded in a standardized, computable form. This work did not directly improve patient care, but it became the semantic foundation that enabled later data-driven and AI approaches. Without the structured terminologies and ontologies built by this framework, the aggregation and analysis of clinical data across institutions would have remained impossible.
By the late 1980s, the proliferation of different EHR systems and clinical databases had created a new problem: data could not flow between them. Clinical Data Standards and Interoperability (1987–Present) emerged as a framework dedicated to solving this communication gap. Standards like HL7 v2 for messaging, DICOM for medical imaging, and later FHIR for modern web-based exchange provided the technical agreements that allowed disparate systems to share data. This framework did not replace the earlier ones; rather, it provided the essential infrastructure that made the EHR a network resource rather than an isolated repository. The standards turn was a prerequisite for the data-driven approaches that followed, because predictive analytics and evidence-based decision support depend on access to large, integrated datasets.
Around the turn of the millennium, the limitations of rule-based decision support became increasingly apparent. The rules were labor-intensive to create, could not easily adapt to new evidence, and often produced too many alerts, leading to alert fatigue. Two new frameworks emerged in response, each with a different epistemological foundation.
EHR-Based Predictive Analytics (2000–Present) shifted the focus from expert-derived rules to statistical patterns learned from historical data. Instead of asking clinicians to encode their knowledge, this framework used machine learning techniques—logistic regression, decision trees, later neural networks—to identify correlations in large EHR datasets. Early applications included predicting hospital readmission, sepsis onset, or patient deterioration. This approach superseded rule-based systems in tasks where patterns were too complex or too numerous to be captured by hand-coded rules. However, it introduced a new problem: the models were often opaque, making it difficult for clinicians to understand why a prediction was made.
At roughly the same time, Evidence-Based Clinical Decision Support (2001–Present) took a different path. Rather than relying on local data patterns, this framework anchored its recommendations in the published medical literature and systematic reviews. It aimed to bring the rigor of evidence-based medicine into the clinical workflow by linking decision support rules to high-quality research findings. This framework coexisted with predictive analytics, but its proponents argued that recommendations grounded in randomized trials and meta-analyses were more trustworthy than those derived from observational data, which could be biased by confounding. The tension between these two frameworks—local, data-driven patterns versus global, evidence-based guidelines—remains a live disagreement in clinical informatics today.
By the late 2000s, it had become clear that neither predictive models nor evidence-based rules alone could transform care. The Learning Health Systems framework (2007–Present) emerged as an aspirational synthesis, drawing on all three earlier decision-support traditions. A learning health system is one in which data from every patient encounter are captured, analyzed, and fed back into care delivery to continuously improve outcomes. This framework derived from EHR-Based Predictive Analytics, Evidence-Based Clinical Decision Support, and Rule-Based Clinical Decision Support, integrating them into a recursive cycle of data collection, analysis, implementation, and evaluation. The learning health system is not a specific technology but an organizational vision that requires all the earlier frameworks to work together. It remains a leading framework today, guiding initiatives at institutions like the U.S. Agency for Healthcare Research and Quality (AHRQ) and many academic medical centers.
At the same time, the Sociotechnical Clinical Informatics framework (2008–Present) offered a critical corrective to the technical optimism of the earlier approaches. Drawing on research in human-computer interaction and organizational studies, this framework argued that the success of clinical information systems depends as much on people, workflows, and organizational culture as on the technology itself. It challenged the assumption that simply deploying a decision support system or an analytics tool would improve care; instead, it insisted that implementation must be studied as a complex adaptive process. Sociotechnical clinical informatics coexists with the other frameworks, providing a lens for understanding why even well-designed systems sometimes fail in practice. Its core research questions include how clinicians adapt to new tools, how alerts affect decision-making, and how to design systems that support rather than disrupt clinical work.
The most recent framework, Clinical Artificial Intelligence and Machine Learning (2012–Present), has intensified the debates that earlier frameworks left unresolved. Deep learning models can now analyze not only structured EHR data but also unstructured clinical notes, medical images, and genomic sequences, expanding the problem domain far beyond what EHR-Based Predictive Analytics could handle. This framework supersedes the earlier predictive analytics approach by offering more powerful methods, but it also reopens the transparency-versus-performance tension. Rule-based systems were fully interpretable; evidence-based systems were grounded in published research; but many modern AI models are black boxes. Clinicians and regulators are demanding explainability, while AI researchers argue that complex models are necessary for high accuracy.
Clinical AI/ML does not replace the earlier frameworks entirely. It coexists with rule-based and evidence-based decision support, often being used for different tasks: AI for pattern recognition and risk prediction, rules for straightforward alerts and reminders, and evidence-based guidelines for standardized protocols. The learning health system vision now incorporates AI as a key component, while sociotechnical informatics researchers study how AI tools are adopted and how they change the roles of clinicians.
Today, the leading frameworks in clinical informatics agree on several points. First, the EHR is the indispensable foundation; no advanced analytics or decision support is possible without reliable, structured clinical data. Second, interoperability standards are necessary to aggregate data across institutions. Third, the learning health system cycle—capture, analyze, implement, improve—is a widely accepted aspirational model. Fourth, sociotechnical factors must be considered in any implementation.
But significant disagreements remain. The most prominent is the tension between transparency and performance: rule-based and evidence-based approaches prioritize interpretability, while AI/ML prioritizes predictive accuracy. A second disagreement concerns the role of local data versus global evidence: should decision support be tailored to a specific institution's patient population (predictive analytics) or based on broadly applicable clinical trials (evidence-based)? A third debate revolves around automation versus augmentation: should AI systems make autonomous recommendations, or should they always support human judgment? These disagreements are not signs of weakness; they reflect the subfield's maturation and the complexity of the problems it addresses. Clinical informatics today is a pluralistic field in which multiple frameworks coexist, each with its own strengths and limitations, and each contributing to the ongoing effort to turn data and knowledge into better care.