Muneeb Ali, Chief Technology Officer, Eon
These days, AI is everywhere — it finishes our text messages and suggests our next Google search. What makes this kind of convenient, automated support possible are Large Language Models (LLMs) that use a large amount of data to "guess” the next word based on previous patterns.
A good guess is fine for a casual query or conversation. For use in a regulated clinical setting — such as reading a radiology report to analyze the characteristics of an incidental finding — AI tools need precision. Clinical data is hidden in narratives, and if an AI "guesses" wrong about an incidental finding, it creates an immediate burden for the clinician.
For an AI tool to be useful in a regulated healthcare setting, it has to do more than just read; it has to understand. Specifically, it needs:
Generalized AI tools can address some of these requirements but consistently struggle with others.
Eon built a proprietary computational linguistics model that could meet the high standards of trustworthiness required from an AI solution built for clinical use. Initially built for use in incidental lung nodule findings, the model can turn unstructured radiology reports into precise, auditable data that clinicians can use and trust, and is built around how medicine is practiced, not how consumer AI works. It pairs the structured extraction capabilities of a generalized AI tool with deep contextual comprehension, so it understands not just the words contained in a radiology report but how they relate to each other and what they mean in a clinical context.
The model understands standard radiology report sections and only treats current “findings” as actionable, avoiding confusion with historical comparisons or background details.
It reliably recognizes findings, sizes, locations, and descriptors across different synonyms and abbreviations, using an anatomy ontology to place each finding in the correct body structure.
The engine links every attribute — size, morphology, location, and change over time — to the correct nodule and tracks growth or stability across serial exams.
Every output is directly tied to the exact sentence in the radiology report, and the system only extracts what is written, never inventing or inferring undocumented findings.
The same report always produces the same result, supporting reproducible workflows, dependable automation, and reliable quality review.
Outputs are delivered as structured fields (finding type, size, location, characteristics, temporal status, recommendations) that can flow directly into EHRs, registries, and tracking systems.
And all these capabilities together mean it can yield better information from any radiology report without forcing radiologists to change how they dictate. The computational linguistic model operates at greater than 99% precision in incidental lung nodule findings, meaning fewer than one false negative or positive for every 100 radiology reports.
After successfully implementing our computational linguistics engine for pulmonary lung nodules, Eon has further expanded its use to additional disease states. Each condition-specific model is aligned to the language, guidelines, and risk criteria of a specific disease area. Clinical interpretation varies significantly across domains. Our specialized models have been customized to the language and guidelines for specific conditions and apply appropriate logic to extract clinically meaningful context.
Lung
Breast
Cardiovascular
Pancreas
Kidney
Thyroid
Liver
By maintaining high precision and consistent behavior, Eon’s computational linguistics approach supports scale without amplifying review workload. Computational linguistics is deterministic and domain-specific, reflecting an architectural approach shaped by sustained clinical use and operational feedback for high precision and reduced downstream burden.
