- The Truveta Language Model (TLM) extracted angiographic occlusion and functional status outcomes from unstructured clinical notes for patients treated with endovascular coiling of unruptured intracranial aneurysms.
- Among patients with available follow-up, angiographic outcomes were generally stable or improved at 6 and 12 months after treatment.
- Functional outcomes were also generally stable or improved over follow-up, demonstrating the feasibility of using large language models to characterize clinically meaningful outcomes from unstructured EHR data.
This report summarizes our poster presented at ISPE 2026, titled Leveraging Unstructured EHR Notes Using Large Language Models to Unlock Real-World Outcomes in Intracranial Aneurysm.
Real-world data can provide an important complement to evidence generated in clinical trials by describing how treatments are used and what outcomes patients experience in routine clinical practice. However, many clinically meaningful outcomes are not consistently represented in structured electronic health record (EHR) fields. Instead, they may be documented in unstructured notes, operative reports, imaging interpretations, and follow-up assessments. This creates a challenge for researchers who rely primarily on structured data because potentially valuable clinical information can remain difficult to identify and analyze at scale.
Intracranial aneurysm treatment provides a clear example. Endovascular coiling is commonly used to treat intracranial aneurysms, and follow-up after treatment often evaluates both the degree of aneurysm occlusion and the patient’s functional status. The Raymond-Roy Occlusion Classification (RROC) is widely used to characterize angiographic occlusion after coiling, with Class I representing complete occlusion, Class II a residual neck, and Class III a residual aneurysm. Prior studies have shown that these classifications can provide important information about the durability of aneurysm treatment and the likelihood of recurrence. Functional status is often assessed using the modified Rankin Scale (mRS), with a 0-6 scale for patient disability ranging from no disability, slight disability, moderate disability, severe disability, to death.
Because RROC and mRS may be recorded primarily in clinical notes, these outcomes are difficult to study using structured EHR and claims data alone. Large language models (LLMs) offer a potential way to extract clinically meaningful information from unstructured documentation at scale, an emerging application of LLMs in real-world data research. We therefore asked: Can the Truveta Language Model (TLM) be used to extract angiographic occlusion and functional outcomes from unstructured EHR notes to characterize real-world outcomes following endovascular coiling of unruptured intracranial aneurysms?
Methods
We conducted a retrospective observational study using a subset of Truveta Data to identify patients with unruptured intracranial aneurysms who underwent endovascular coiling between 2016 and 2025. Truveta Data combines de-identified structured EHR information with unstructured clinical notes from more than 1340 million patients across the United States.
Patients were eligible for the analysis if they had evidence of an unruptured intracranial aneurysm treated with an endovascular coiling procedure during the study period. The final cohort included 8,925 patients.
We used the Truveta Language Model (TLM) to extract two clinically relevant outcomes from unstructured EHR documentation: angiographic occlusion using the Raymond-Roy Occlusion Classification (RROC, Classes I–III) and functional status using the modified Rankin Scale (mRS, Levels 0–6). Outcomes were characterized at approximately 6 and 12 months following the index coiling procedure.
We used descriptive analyses to summarize patient characteristics and the distribution of RROC and mRS outcomes over follow-up. We also examined changes in outcome classifications relative to baseline among patients with available follow-up data.
Results
Patient population
The study included 8,925 patients who underwent endovascular coiling for an unruptured intracranial aneurysm between 2016 and 2025. The mean age was 61.2 years (SD 13.3), 73.0% of patients were female, and 38.4% were current or former tobacco users. Mean baseline systolic blood pressure was 139.4 mm Hg (SD 27.6), while mean baseline diastolic blood pressure was 79.3 mm Hg (SD 15.5).
TLM-extracted follow-up outcomes were available for a subset of the overall cohort. RROC information was available for 1,533 patients at 6 months and 1,235 patients at 12 months. mRS information was available for 1,400 patients at 6 months and 758 patients at 12 months. Thus, while the overall cohort was large, longitudinal outcome documentation was available for a smaller proportion of patients.
Angiographic outcomes following coiling
Among patients with available RROC follow-up, angiographic outcomes were generally stable or improved following treatment. At 6 months, 33.8% of patients with a baseline Class II occlusion and 37.8% of those with a baseline Class III occlusion demonstrated stability or improvement. Most patients with baseline Class I occlusion remained stable.
These findings are consistent with previous research showing that incompletely occluded aneurysms do not necessarily progress after coiling and that some can remain stable or progress toward more complete occlusion over time. Prior studies have also demonstrated that the type of residual filling captured by the Raymond-Roy classification can provide information about subsequent aneurysm durability.
Functional outcomes following coiling
Functional outcomes assessed using the mRS were also generally stable or improved during follow-up. Among patients with a 6-month follow-up score, most patients with baseline mRS scores of 0 or 1 remained at the same level, indicating preserved functional status. Similar patterns were observed among patients with higher baseline mRS scores.
The consistency of these patterns demonstrates that TLM was able to identify and characterize a clinically meaningful functional outcome that may otherwise be difficult to capture from structured EHR data alone.
Discussion
In this large real-world cohort of 8,925 patients treated with endovascular coiling for unruptured intracranial aneurysms, angiographic occlusion and functional outcomes were generally stable or improved among patients with available longitudinal documentation. More importantly, the study demonstrates a broader capability: clinically meaningful outcomes that are primarily documented in unstructured notes can be transformed into analyzable real-world data using the Truveta Language Model.
For intracranial aneurysm treatment, RROC is an established framework for evaluating the degree of aneurysm occlusion after coiling, and prior work has shown that changes in occlusion classification can provide insight into treatment durability and recurrence risk. Similarly, the mRS is a widely used measure of functional outcome, although its application can be affected by variability in clinical assessment. Being able to identify these outcomes directly from clinical documentation creates an opportunity to study treatment effectiveness using data that would otherwise be difficult to analyze at scale.
The broader implication extends beyond intracranial aneurysms. Structured EHR data are well suited to capturing diagnoses, procedures, laboratory values, and medications, but many outcomes that matter to patients and clinicians are expressed in note form. Symptoms, disease severity, treatment response, imaging findings, functional status, adverse events, and clinical assessments may all be documented in notes. Recent research has demonstrated the potential of LLMs to extract structured clinical information from unstructured documentation, supporting the use of these approaches for large-scale secondary research.
TLM therefore has the potential to expand the range of questions that can be answered using real-world data. Rather than limiting research to the information that happens to be available in structured fields, researchers can use clinical narratives as an additional source of longitudinal evidence. Across therapeutic areas, this could enable more detailed characterization of treatment outcomes, disease progression, patient characteristics, and clinical events.
Several limitations should be considered when interpreting these findings. First, longitudinal follow-up was limited: RROC data were available for 39.2% of patients at 6 months and 25.2% at 12 months, while mRS data were available for 1,400 patients at 6 months and 758 at 12 months. Patients with documented follow-up may therefore differ from those without available follow-up, and the observed outcome patterns should not be interpreted as representative of the full treated population. Second, the analysis relied on information documented in clinical notes. Missing or inconsistently documented assessments may have limited the availability of outcomes. Finally, this was a descriptive study and was not designed to establish causal relationships between treatment and clinical outcomes.
Despite these limitations, this analysis demonstrates the potential of combining structured EHR data with unstructured clinical notes through TLM. By turning unstructured documentation into research-ready information, TLM can help unlock outcomes that have historically been difficult to study at scale. The intracranial aneurysm analysis is one example of a broader opportunity to use real-world clinical narratives to generate evidence across therapeutic areas.
Data are constantly changing and updating. These findings are consistent with data analyzed for the ISPE 2026 study.
Citations
- Mascitelli RJ, Moyle H, Oermann EK, et al. An update to the Raymond-Roy Occlusion Classification of intracranial aneurysms treated with coil embolization. J Neurointerv Surg. 2015;7:496–502.
- Cloft HJ, Kallmes DF. Aneurysm recurrence after endovascular treatment. [Reference to be finalized based on the team’s preferred background source.]
- van Swieten JC, Koudstaal PJ, Visser MC, Schouten HJA, van Gijn J. Interobserver agreement for the assessment of handicap in stroke patients. Stroke. 1988;19:604–607.
- Quinn TJ, Dawson J, Walters MR, Lees KR. Exploring the reliability of the modified Rankin Scale. Stroke. 2009;40:762–766.
- Hynes DM, et al. Outcomes validity and reliability of the modified Rankin Scale: implications for stroke clinical trials.
