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

- **DOI:** [10.1136/jnnp-2025-337689](https://doi.org/10.1136/jnnp-2025-337689)
- **Rating:** Sound (3.5 out of 5)
- **Verdict:** Whole-population NLP of routine brain imaging turns incidental covert cerebrovascular findings into credible stroke and dementia risk markers, undercut by causal framing the observational design cannot carry.
- **Framework version:** v5.9
- **Pipeline version:** v0.4
- **Evaluation date:** 2026-10-08
- **Evaluation page:** https://www.nabu.science/paper/10.1136%2Fjnnp-2025-337689

## Key claims

1. CCD and atrophy were associated with a higher risk of both dementia and stroke up to 12 years later, with important differences between CCD exposures.
   - Support: Supported
   - Type: Primary finding
   - Basis: Large linked cohort with lag and exclusion analyses, but unmeasured lifestyle confounding and unclassifiable imaging indication limit identification.
   - Location: Discussion, final paragraph (Conclusion)
2. We demonstrate a higher risk of dementia (particularly Alzheimer's disease) in people with cerebral atrophy and a higher risk of stroke in people with cortical infarcts, but no significant associations with an age-associated control outcome (colorectal cancer), supporting a causal relationship.
   - Support: Tentative
   - Type: Secondary
   - Basis: One non-brain negative control sharing confounders cannot support causation in a design conceding residual confounding.
   - Location: Box 'What this study adds'
3. CCD and atrophy have implications for future disease risk and can be identified at scale using NLP of clinical reports.
   - Support: Supported
   - Type: Secondary
   - Basis: Pipeline was run across an entire national population; phenotype accuracy rests on validation reported outside this paper.
   - Location: Abstract, Conclusion
4. Compared with those with no CCD or atrophy phenotype, those with each phenotype (accounting for the presence of other phenotypes) had a higher risk of any stroke over the 12-year follow-up period.
   - Support: Supported
   - Type: Secondary
   - Basis: Mutually adjusted hazard ratios with tight intervals across all four phenotypes, though atrophy's estimate is marginal and detection bias is unquantified.
   - Location: Results §Risk of stroke; Figure 3

## Strengths

- **Whole-population cohort with multi-phenotype NLP** All 367,988 people in Scotland with brain imaging and no prior neurological disease were included, with four phenotypes extracted from both CT and MRI reports generated in routine care. This covers district hospitals as well as academic centres, a population that research imaging cohorts systematically miss.
  - Dimension: contribution
  - Location: Methods §Study design and participants; Figure 1; Discussion, strengths paragraph
- **Reverse-causation safeguards built into the design** Follow-up begins six months after the index scan, people diagnosed within that window are excluded, and sensitivity analyses repeat the models excluding the first 1 and 5 years. Where estimates shift, the paper reports it openly, noting that the atrophy-stroke association moves to the null at 5 years.
  - Dimension: methodological rigour
  - Location: Methods, follow-up paragraph; Results §Sensitivity analyses
- **Negative controls and competing-risk accounting** Epilepsy and colorectal cancer are prespecified as brain-based and non-brain control outcomes, and the cumulative incidence estimates account for the competing risk of death in an elderly, high-mortality cohort. Proportional hazards were tested and violating variables handled by stratification.
  - Dimension: methodological rigour
  - Location: Methods §Study definitions and analysis; Table 2; Results §Risk of Parkinson's disease, epilepsy and colorectal cancer
- **Candid, specific and traceable limitations** The Discussion names the rules-based rather than LLM approach, clinical selection, unclassifiable imaging indication, missing primary-care and lifestyle data, dementia-subtype diagnosis bias, epilepsy's imperfection as a control and the crudeness of the phenotypes, each linked to its effect on interpretation.
  - Dimension: positioning
  - Location: Discussion, limitations paragraph

## Limitations

- **Causal framing exceeds the observational design** The summary box concludes that the absence of a colorectal-cancer association supports 'a causal relationship'. A single non-brain negative control, which shares confounders such as smoking with the exposures, cannot establish causation in a design whose own limitations concede that residual confounding 'could explain some of our findings'.
  - Dimension: reporting
  - Location: Box 'What this study adds'; Discussion, limitations paragraph
- **Unmeasured confounding left unquantified** Smoking, obesity, blood pressure and primary-care records are unavailable and the exposed group differs sharply from the unexposed on age and comorbidity (mean 73.4 vs 48.2 years). No quantitative bias analysis, E-value or causal diagram is offered to bound how much of the association this could explain.
  - Dimension: methodological rigour
  - Location: Results, Table 1; Discussion, limitations paragraph
- **Population-comparison relative risks are not like-for-like** The analytic cohort excludes prior disease and requires six-month survival while the wider Scottish comparator does not, producing figures such as a 5-year stroke RR of 30.3. The same comparison yields a 5-year colorectal cancer RR of 5.1 (4.9-5.3) while the text describes the group as not more likely to experience colorectal cancer, which clouds the negative control's interpretation.
  - Dimension: methodological rigour
  - Location: Results, paragraph on comparison with the wider Scottish population
- **Inverse Parkinson's finding reported but unexplained** Cortical infarct is associated with a lower risk of Parkinson's disease (aHR 0.7, 95% CI 0.6 to 0.9), a direction opposite to the paper's framing. Competing mortality risk and selection are plausible mechanisms, but the Discussion does not engage the result.
  - Dimension: positioning
  - Location: Results §Risk of Parkinson's disease; Abstract
- **Multiplicity uncorrected across many tests** Dozens of phenotype-outcome and subgroup interaction tests are reported with p values explicitly uncorrected for multiple comparisons. In a cohort of this size, statistically significant interactions of limited substantive importance are expected.
  - Dimension: methodological rigour
  - Location: Methods, final analysis paragraph; Results §Subgroup analyses

## Reliability checks

Caveats (4 of 4 checks).

- **Results check:** Noted
  - CCD and atrophy were associated with a higher risk of both dementia and stroke up to 12 years later, with important differences between CCD exposures: "The aHR of dementia was WMH 1.3 (1.3-1.3), lacunes 1.0 (0.9-1.0)" (Abstract, Results) vs "although not lacunes (1.0, 95% CI 1.0 to 1.0)" (Results, Risk of dementia). The same lacunes–dementia estimate has a lower confidence limit of 0.9 in the abstract and 1.0 in Results, and the supplied text does not establish which is correct. → Effect on result unclear
  - Minor reporting notes (2):
    - Descriptive: "Incidence was proportionally higher in those with a brain image report than in the wider population for all outcomes except colorectal cancer, though RRs were lower with increasing age (online supplemental material)." (Results, comparison with the wider Scottish population) vs "but not colorectal cancer (RR 1.1, 95% CI 0.9 to 1.2; RR 5.1, 95% CI 4.9 to 5.3)." (Results, comparison with the wider Scottish population). The reported five-year colorectal cancer RR of 5.1 has a confidence interval entirely above 1, contradicting the statement that colorectal cancer risk was not elevated at both reported time points. → Not supported by the paper's own data
    - Descriptive: "The aHR of Parkinson's disease was WMH 1.1 (1.0-1.2), lacunes 1.1 (0.9-1.2)" (Abstract, Results) vs "but not clearly higher in people with lacunes (1.1, 95% CI 1.0 to 1.2)" (Results, Risk of Parkinson’s disease, epilepsy and colorectal cancer). The same lacunes–Parkinson’s disease estimate has a lower confidence limit of 0.9 in the abstract and 1.0 in Results, and the supplied text does not establish which is correct. → Effect on result unclear
- **Reporting and transparency:** Clean. No reporting or transparency concern can be retained with the required textual grounding; the stated restricted data availability is consistent with the described safe-haven analysis.
  - Ethics approval: declared · Data availability: declared
- **Reference integrity:** Clean
  - 35 references in manuscript
  - 29 of 29 checkable references found in an index
  - 6 are books, websites or datasets — counted, but not index-checkable
  - References not verified automatically are resolved manually within 24 hours.

## Post-publication record

- **Status:** Clean
- **Summary:** No retraction or expression of concern found in Crossref (includes Retraction Watch).
- **Last checked:** 2026-10-17
