Semantic Strategy
& Architecture
for the AI Era

Modern AI changes two key activities pertinent to e-health:

• software economics: it radically reduces the cost and time to produce software and applications;
• insight generation: it provides new ability to extract insights from data, including differential diagnoses, risk reports, as well as population level analytics.

Behind both lies the central LLM capability of text prediction. This provides a means of translating data across formats, languages and standards.

Key limitations of LLM AI based environments (not necessarily LLM itself) include:

• LLM AI functions at the language level, not on semantics, and routinely creates semantic errors in its output (underlying problems: heterogeneous source data; non-semantic tokenisation);
• it has no way on its own of distinguishing fact from fiction (underlying problems: no inbuilt reasoner, ontology, or reliable training set);
• missing real world context, including older data (e.g. patient problems, social situation), tacit knowledge (e.g. type of clinic), unrecorded knowledge (e.g. patient preferences);
foundation model training bases include false information, unscientific biases and time-limited claims.

The first problem can be understood as the ‘straight from text’ problem: in an agentic AI system, source data is accessed as syntax, and goes straight to a black box, which generates output. It is impossible to know what semantics are assumed of the data, or see any reasoning process. The second problem is the lack of connection of semantics to the syntax. The third problem has to do with the partial nature of the data originally accessed, allowing inappropriate conclusions to be generated, and possibly causing patient harm. For many mundane tasks in medicine, such as assistance with writing notes from boilerplate, or summarising large medical texts, these problems will not manifest. For decision support, cohort identification and analytics, they will. However, like fatigue cracking in airframes, they will not initially be visible.

There are sociological risks of AI in healthcare as well: over-reliance by professionals on automated systems (as for pilots on fly-by-wire) may lead to scenarios in which professionals do not know how to proceed on their own.

If AI is to create value in healthcare without creating patient risk or atrophying professional skill levels, new strategies must be developed that account for semantics, traceability, and new modes of training.

The combination of multi-domain application, semantic computability and realtime domain modelling in SPLASH creates a data platform that tracks the domain rather than becoming obsolete, and greatly improves the quality of Agentic AI, which otherwise is limited to numerous highly variable linguistic representations, no standardisation of terms, and no access to ontologic truths.

With the SPLASH platform, decision support and insight generation improve markedly, while hallucination and context-unaware errors are greatly reduced. Many of the innovations of SPLASH can be realised in the current generation of openEHR systems.

Without common upstream domain models, or a common ontology foundation layer, standards artifacts are guaranteed to create more variation, due to localised and vendor-specific structuring of data and terminology use. This actually worsens the problem, while not addressing the need for a common semantics (i.e. a ‘common language’ of healthcare).

However, with appropriate tools, profiles can be generated from common upstream models, ensuring that even privately created forms still conform to the common semantics.