Prognosis of stroke subtypes in whole population health systems data: a matched cohort study

By applying natural language processing to brain imaging reports linked with nationwide health data, this matched cohort study successfully subtyped stroke cases in Scotland to reveal distinct risks of death, rehospitalization, and comorbidities across different stroke subtypes.

The Gist

Imagine the human brain as a vast, complex city. When a "stroke" happens, it's like a major traffic accident or a power outage in a specific neighborhood of that city. For a long time, doctors and researchers have known that where the accident happens matters just as much as what happened. A crash in the downtown business district (the cortex) causes different problems than a crash in the deep underground tunnels (the deep brain structures).

However, there was a huge problem: The city's official logbooks (hospital records) were too vague. They would just write "Traffic Accident" without saying which neighborhood was hit. This made it impossible to study the long-term consequences of specific types of accidents on a large scale.

This paper is like a team of super-smart detectives who decided to fix those logbooks by reading the handwritten notes of the city's engineers (radiologists).

The Detective Work: Reading Between the Lines

The researchers in Scotland had access to millions of brain scan reports. While the official computer codes were vague, the radiologists who wrote the reports used plain English to describe exactly what they saw.

Instead of hiring thousands of people to read every single report, the team built a digital detective robot (called Natural Language Processing, or NLP). This robot was trained to read the free-text notes and instantly spot the difference between:

• Ischemic Stroke: A clogged pipe (blockage).
• Hemorrhage: A burst pipe (bleeding).
• Location: Whether the damage was in the "downtown" (cortical/lobar) or the "underground tunnels" (deep).

By using this robot, they turned a messy pile of vague records into a crystal-clear map of 64,000 specific stroke events.

The Big Findings: What Happens After the Crash?

Once they had this clear map, they compared the people who had these specific "accidents" against a group of people who never had a stroke (the control group). Here is what they discovered, using some simple analogies:

1. The "Burst Pipe" in the Downtown (Lobar ICH)

• The Finding: People who had bleeding in the outer layer of the brain (lobar) were much more likely to develop dementia later in life compared to those with bleeding deep inside.
• The Analogy: Imagine a burst pipe in the fancy, high-end apartments (the cortex). Even after the water is mopped up, the walls are stained, and the building's structure is weakened. Over time, this makes the whole building (the brain) much more likely to fall apart (dementia) than if the burst pipe had happened in the sturdy, concrete basement (deep brain).

2. The "Clogged Pipe" in the Downtown (Cortical Ischemic Stroke)

• The Finding: People with blockages in the outer brain were at a much higher risk of having a heart attack (Myocardial Infarction) in the months immediately following their stroke.
• The Analogy: If the "downtown" traffic jam was caused by a clogged pipe, it suggests the entire plumbing system in the city is old and rusty. The same rust that clogged the brain pipe is likely clogging the heart pipes too. So, shortly after the brain crash, the heart is also at high risk of failing.

3. The "Deep Tunnel" vs. "Downtown" Seizures

• The Finding: Seizures (epilepsy) were much more common after strokes in the outer layers (cortical/lobar) than in the deep layers.
• The Analogy: The outer layer of the brain is like the surface of a lake; it's very sensitive to ripples. A crash here sends shockwaves that easily trigger electrical storms (seizures). The deep tunnels are more insulated; a crash there is contained and less likely to cause a storm on the surface.

4. The Immediate Danger

• The Finding: Bleeding strokes (Hemorrhagic) were far more deadly in the first six months than blockage strokes (Ischemic).
• The Analogy: A burst pipe (bleeding) causes immediate, catastrophic flooding that can drown the building quickly. A clogged pipe (blockage) is a slow suffocation; it's dangerous, but the building has more time to adapt before it collapses.

Why This Matters

Before this study, researchers were trying to solve a puzzle with half the pieces missing. They knew strokes were bad, but they didn't know which strokes led to which specific future problems.

By using their "digital detective" to read the handwritten notes, they filled in the missing pieces. Now, doctors can tell a patient:

"Because your stroke happened in the 'downtown' area, we need to watch your heart closely for the next six months."
"Because your bleeding was in the outer layer, we should start planning for memory support sooner."

The Bottom Line

This paper shows that we don't need to throw away our old, messy data. We just need the right tools (like AI and smart algorithms) to read the stories hidden inside the text. This allows us to predict the future of stroke survivors with much greater accuracy, helping doctors provide better, more personalized care for everyone.

Technical Summary

1. Problem Statement

Traditional studies on stroke outcomes are often limited by small cohort sizes, selection bias, and short follow-up periods. Furthermore, large-scale Healthcare Systems Data (HSD) studies typically rely on standardized coding ontologies (e.g., ICD-10) to identify stroke subtypes. These codes lack granularity; they often fail to distinguish between ischemic and hemorrhagic stroke, and crucially, they do not specify the anatomical location (e.g., cortical vs. deep, lobar vs. deep) of the stroke. This lack of phenotypic detail prevents researchers from accurately estimating the specific risks of long-term outcomes (such as dementia, epilepsy, or myocardial infarction) associated with distinct stroke subtypes.

2. Methodology

The study employed a matched cohort design using nationwide, linked health data from Scotland (2010–2018) to overcome the limitations of coded data.

• Data Sources: The researchers linked Scottish Medical Imaging data (CT and MRI head scan reports) with routine administrative data, including:
  • Hospital admissions (SMR01/04).
  • Cancer registry (SMR06).
  • Community prescriptions (Prescribing Information System).
  • Death records (NRS Scotland).
• Natural Language Processing (NLP):
  • They applied a previously validated, rules-based NLP algorithm (EdIE-R) to free-text radiology reports.
  • The algorithm extracted specific phenotypes: Ischemic Stroke, Intracerebral Hemorrhage (ICH), Subarachnoid Hemorrhage (SAH), and Subdural Hemorrhage (SDH).
  • Crucially, it subclassified these by location: Cortical vs. Deep (for ischemic) and Lobar vs. Deep (for ICH).
  • It also tagged comorbidities like brain atrophy and white matter hyperintensities (WMH).
• Cohort Definition:
  • Cases: Identified 64,219 individuals with clinical stroke phenotypes confirmed by both ICD-10 codes and NLP tags.
  • Controls: For every case, four controls without a history of stroke were selected, matched by age (±2 years) and sex.
  • Exclusions: Patients with head trauma or brain cancer preceding the hemorrhage were excluded to ensure vascular etiology.
• Statistical Analysis:
  • Outcomes: Rehospitalization (stroke, MI, cancer, dementia, epilepsy) and death.
  • Models: Cause-specific Cox proportional hazard models were used to estimate Adjusted Hazard Ratios (aHR).
  • Time Splits: Due to non-proportional hazards, analysis was split into two periods: <6 months** (early) and **>6 months (late) post-stroke.
  • Confounders: Adjusted for age, sex, diabetes, hypertension, atrial fibrillation, and medication usage.

3. Key Contributions

• Scale and Granularity: Successfully subtyped stroke at an unprecedented scale (64,219 cases) using NLP on free-text reports, reducing the proportion of "unspecified" stroke diagnoses from 26.1% to 3.4%.
• Methodological Validation: Demonstrated that linking NLP-extracted imaging data with routine coded HSD is a viable, scalable method for high-resolution epidemiology without requiring manual chart review.
• Phenotype-Specific Risk Profiling: Provided the first large-scale, population-level estimates of long-term risks (dementia, epilepsy, MI) specifically stratified by stroke location (cortical/deep, lobar/deep).

4. Key Results

The study identified distinct risk profiles based on stroke subtype and location:

• Stroke Recurrence:
  • Lobar ICH had a significantly higher risk of readmission for stroke compared to Deep ICH (aHR 1.71).
  • No significant difference was found between cortical and deep ischemic stroke regarding recurrence.
• Myocardial Infarction (MI):
  • Cortical Ischemic Stroke was associated with the highest risk of early MI (<6 months) compared to other subtypes (aHR 4.6). This suggests a shared pathological origin (large artery atherosclerosis) between cortical stroke and cardiac events.
• Dementia:
  • Lobar ICH carried the highest long-term risk of dementia (>6 months) compared to controls (aHR 3.5) and other subtypes.
  • Cumulative incidence of dementia by 5 years in men >70 was 17% for lobar ICH vs. ~10% for other subtypes.
• Epilepsy:
  • Risk was elevated for Cortical Ischemic and Lobar ICH subtypes compared to their deep counterparts, both early and late post-stroke.
• Mortality:
  • ICH had a much higher early mortality rate than ischemic stroke (45.8% died within 6 months vs. 21.7%).
  • Cortical Ischemic Stroke had a higher risk of death compared to Deep Ischemic Stroke in both the short and long term.
• Cancer:
  • A modestly elevated cancer risk was observed in the first 6 months for some subtypes, but no long-term increased risk was detected compared to controls.

5. Significance and Implications

• Clinical Relevance: The findings validate that stroke location is a critical prognostic factor. For instance, patients with lobar ICH require specific monitoring for dementia, while those with cortical ischemic stroke need heightened cardiac surveillance.
• Healthcare Systems: This approach proves that "uncoded" free-text data (radiology reports) can be mined at scale to improve the quality of epidemiological research and health system audits.
• Future Applications: The methodology can be extended to other conditions where imaging reports contain critical phenotypic details not captured in standard coding. It also offers a pathway to identify participants for clinical trials within large national cohorts (e.g., UK Biobank) by automating phenotype extraction.
• Limitations: The study relied on a rules-based NLP system rather than modern Large Language Models (LLMs) due to governance constraints in secure data environments. Additionally, roughly half of the ICH and ischemic cases could not be subtyped by location, though their risk profiles were similar to those with identified locations.

In conclusion, this study successfully bridges the gap between high-resolution clinical phenotyping and large-scale population data, providing robust evidence that stroke prognosis is heavily dependent on anatomical subtype.