Clinically reported covert cerebrovascular disease and risk of neurological disease: a whole-population cohort of 367,988 people using natural language processing

Using natural language processing on brain imaging reports from a Scottish cohort of nearly 370,000 people, this study demonstrates that clinically reported covert cerebrovascular disease and cerebral atrophy are significant predictors of future stroke, dementia, and Parkinson's disease, highlighting the need for targeted prevention strategies.

The Gist

Imagine your brain is like a bustling city. Sometimes, even when the city looks fine from the outside and everyone is going about their daily business, there are hidden cracks in the roads, potholes in the pavement, or signs of wear and tear on the buildings. In the medical world, these hidden issues are called Covert Cerebrovascular Disease (CCD). They are "covert" because the person doesn't feel sick or have symptoms like a stroke, but the damage is there.

For a long time, doctors knew these cracks existed, but they didn't have a good way to count them in the general population or predict exactly how much trouble they might cause later. It was like trying to guess how many potholes are in a whole country just by looking at a few photos.

The Big Detective Hunt

This study was a massive detective mission involving 367,988 people in Scotland. The researchers wanted to answer a simple question: If a brain scan shows these hidden cracks, how likely is that person to have a major accident (like a stroke or dementia) in the future?

The Problem: Most medical records are written in plain English, not in neat checkboxes. A radiologist might write, "There are some white spots on the brain," or "The brain looks a bit shrunken." Computers usually can't read these sentences to find patterns.

The Solution: The team used a special computer program called Natural Language Processing (NLP). Think of this NLP as a super-smart, tireless librarian who can read thousands of messy, handwritten-style reports in seconds. It was trained to spot specific phrases that meant "hidden damage" and categorize them into four types:

1. White Matter Spots (WMH): Like small patches of rust on the city's pipes.
2. Lacunes: Tiny, deep potholes.
3. Cortical Infarcts: Larger cracks near the surface of the city.
4. Atrophy: The city itself looking a bit smaller or shrunken (like a fruit drying out).

What They Found

The librarian found that 35% of the people scanned had at least one of these hidden issues, even though they hadn't been diagnosed with a brain disease yet. Here is what the data revealed about their future risks:

• The "Big Crash" (Stroke): If the scan showed Cortical Infarcts (the large surface cracks), the risk of having a stroke later was the highest. It's like finding a major crack in a bridge; you know a collapse is more likely.
• The "Slow Fade" (Dementia): If the scan showed Atrophy (shrinkage), the risk of developing dementia was the highest. It's like a city slowly losing its population and infrastructure; the decline is gradual but significant.
• The "Small Potholes" (Lacunes): These were linked to a higher risk of bleeding strokes (hemorrhagic).
• The "Rust" (White Matter Spots): These were linked to a higher risk of vascular dementia (dementia caused by blood flow issues).

The "Control" Test: To make sure the computer wasn't just finding random patterns, the researchers checked if these brain cracks were linked to colorectal cancer (a disease of the colon). They weren't. This is like checking if a pothole in the road causes a broken leg in a different city. Since there was no link, it proved that the brain cracks were specifically linked to brain problems, not just general bad luck.

The Takeaway for Everyday Life

This study is a game-changer because it proves that hidden brain damage is a warning sign, even if you feel fine today.

1. It's a "Check Engine" Light: Finding these issues on a routine scan is like seeing a "Check Engine" light on your car dashboard. You aren't broken down yet, but you need to pay attention.
2. Different Cracks, Different Risks: Not all brain damage is the same. A shrunken brain points to dementia risks, while surface cracks point to stroke risks. Doctors can now look at a report and say, "Okay, this specific type of finding means we should focus on preventing strokes," or "This one means we need to watch for memory issues."
3. Prevention is Key: Since these people don't feel sick, they might not be getting treatment. The study suggests that if a doctor sees these "covert" signs, they should be extra aggressive about managing blood pressure and cholesterol to prevent the "big crash" later.

In short: This research used a digital "super-reader" to find hidden brain damage in hundreds of thousands of people. It confirmed that these silent signs are real predictors of future brain trouble. By catching them early, doctors and patients can take steps to keep the "city" running smoothly for longer.

Technical Summary

1. Problem Statement

Covert Cerebrovascular Disease (CCD) refers to brain imaging findings (such as white matter hyperintensities, lacunes, and cortical infarcts) in individuals who have never experienced symptomatic stroke or transient ischemic attack (TIA). While systematic reviews of research-based MRI cohorts suggest CCD increases the risk of future stroke, dementia, and Parkinson's disease, there is a critical gap in understanding its relevance in routine clinical practice.

• Limitations of existing data: Most evidence comes from research cohorts that are not representative of the general clinical population. Furthermore, CCD lacks a specific diagnostic code in standard classifications (e.g., ICD-11), making it difficult to identify in large-scale electronic health records (EHRs) using traditional coding methods.
• Clinical Uncertainty: It is unclear whether clinically reported CCD findings on routine CT or MRI scans (which are less standardized than research scans) predict future neurological disease in a real-world setting.

2. Methodology

The study employed a whole-population cohort design utilizing linked health data from Scotland.

• Data Source & Cohort:

  • Population: 367,988 individuals from Scotland who underwent brain imaging (MRI or CT) between 2010 and 2018.
  • Inclusion Criteria: Individuals with no prior hospitalization for neurological disease (stroke, dementia, epilepsy, Parkinson's, etc.) before the scan or within 6 months after.
  • Exclusion: Those with missing demographics, conflicting dates, research-only reports, or non-brain imaging.
  • Follow-up: Up to 12 years (until Dec 31, 2020), starting 6 months post-scan to allow for diagnosis latency.

• Natural Language Processing (NLP):

  • Tool: A validated, rules-based NLP algorithm (EdIE-R) was applied to unstructured radiology reports.
  • Phenotypes Identified: The algorithm extracted four specific imaging phenotypes:
    1. White Matter Hypoattenuation/Hyperintensities (WMH).
    2. Lacunes (small subcortical infarcts).
    3. Cortical Infarcts.
    4. Cerebral Atrophy.
  • Validation: The tool was previously validated against expert reads of Scottish imaging reports and underlying images.

• Outcomes & Controls:

  • Primary Outcomes: Incident Stroke (subtypes: ischemic, hemorrhagic, uncertain), Dementia (subtypes: Alzheimer's, vascular, unspecified), and Parkinson's Disease.
  • Negative Controls: Epilepsy (brain-based but not directly linked to CCD) and Colorectal Cancer (non-brain condition) to test for residual confounding.

• Statistical Analysis:

  • Models: Cox proportional hazards models to calculate adjusted Hazard Ratios (aHR).
  • Adjustments: Models were maximally adjusted for age, sex, deprivation (SIMD), rurality, scan modality (CT vs. MRI), pre-scan healthcare usage, and comorbidities (MI, heart failure, diabetes, etc.).
  • Sensitivity: Analyses excluded the first 1 and 5 years of follow-up to test for reverse causality (undiagnosed disease at baseline).

3. Key Contributions

• Scale and Scope: This is one of the largest studies of its kind, analyzing over 367,000 clinical patients (significantly larger than typical research cohorts like UK Biobank) across both academic and district hospitals.
• NLP Application: Demonstrated the feasibility and accuracy of using NLP to extract complex, non-coded imaging phenotypes from routine clinical reports at a population scale.
• Phenotype Differentiation: Unlike many studies that group CCD findings, this study distinguished between specific phenotypes (WMH, lacunes, cortical infarcts, atrophy) and their distinct associations with different neurological outcomes.
• Control Validation: The inclusion of negative controls (colorectal cancer) showed no association, strengthening the causal inference that the observed risks are specific to brain pathology rather than general health deterioration.

4. Key Results

From the cohort, 35% (129,199) had at least one CCD or atrophy phenotype.

Risk of Stroke:

• All phenotypes were associated with increased stroke risk.
• Cortical Infarcts showed the highest risk (aHR 1.8).
• Lacunes (aHR 1.6) and WMH (aHR 1.4) also showed significant risks.
• Cerebral Atrophy had a modest association (aHR 1.1).
• Subtypes: Cortical infarcts were most strongly linked to ischemic stroke; lacunes and cortical infarcts were linked to intracerebral hemorrhage.

Risk of Dementia:

• Cerebral Atrophy was the strongest predictor of all dementia types (aHR 1.7), particularly Alzheimer's disease (aHR 1.9).
• WMH was strongly associated with vascular dementia (aHR 2.0).
• Cortical Infarcts and Lacunes were associated with vascular dementia but showed a reduced risk of Alzheimer's disease (aHR ~0.7).

Risk of Parkinson's Disease:

• Cerebral Atrophy (aHR 1.4) and WMH (aHR 1.1) were associated with increased risk.
• Cortical Infarcts were associated with a lower risk (aHR 0.7).

Control Outcomes:

• No significant associations were found between any CCD phenotype and epilepsy or colorectal cancer, supporting the specificity of the findings.

Subgroup Findings:

• Associations were generally stronger in younger patients compared to older patients.
• MRI scans showed stronger associations for lacunes and atrophy compared to CT scans.
• There were sex-specific differences (e.g., WMH-stroke risk was higher in males; atrophy-stroke risk was higher in females).

5. Significance and Implications

• Clinical Relevance: The study confirms that "incidental" findings on routine clinical brain scans are potent risk markers for future stroke and dementia. A report of atrophy suggests a high risk of dementia, while cortical infarcts/lacunes suggest a high risk of stroke.
• Preventive Strategy: Since CCD is currently uncoded and often ignored in routine care, these findings suggest a need for targeted primary prevention (e.g., aggressive management of hypertension and lipids) in patients with these imaging findings, even if they are asymptomatic.
• Methodological Advance: The study validates NLP as a scalable tool for epidemiological research, allowing researchers to leverage vast amounts of unstructured clinical text that were previously inaccessible.
• Policy Impact: The authors call for clinical trials to test interventions specifically designed to reduce neurological disease risk in patients identified with CCD via routine imaging.

Limitations Noted:

• Use of rules-based NLP rather than Large Language Models (LLMs), though performance was high.
• Selection bias (clinical population is sicker than the general population).
• Lack of primary care data (e.g., smoking, obesity) leading to potential residual confounding.
• Inability to quantify the severity or precise location of lesions due to the nature of clinical reports.