Diagnostic Transportability: What the Criteria Can and Cannot Do

Why a diagnostic criterion has no context-free accuracy

Sensitivity and specificity are properties of a test in a particular spectrum of disease, acquisition conditions, reference standard, and interpretation process. Positive and negative predictive values additionally depend on prevalence. A criterion derived among symptomatic patients selected for MRI and tissue is answering a different question from one applied to an asymptomatic community sample.

This is not a technical footnote. It is one of the most important current lessons in CAA diagnosis.

Boston criteria v2.0 in the derivation setting

The 2022 multicentre study included adults aged 50 years or older with spontaneous intracerebral hemorrhage, cognitive impairment, or transient focal neurologic episodes, along with adequate MRI and histopathologic evaluation. Probable CAA required either at least two strictly lobar hemorrhagic lesions or one strictly lobar hemorrhagic lesion plus severe centrum-semiovale perivascular spaces or a multispot white-matter hyperintensity pattern, with no deep hemorrhagic lesion and no better alternative cause.

Performance was strong within those sampled clinical settings. Sensitivity and specificity were 74.8% and 84.6% in derivation, 92.5% and 89.5% in temporal validation, and 80.2% and 81.5% in geographic validation. In the full-autopsy subset they were 74.5% and 95.0%, respectively.

Those numbers support Boston v2.0 for the intended symptomatic clinical spectrum. They do not establish equivalent screening performance.

Community and nonhemorrhagic validation

In 134 ROSMAP participants with in vivo 3T MRI and autopsy, probable CAA v2.0 sensitivity was 38.8% and specificity 83.5%. The updated white-matter features increased sensitivity relative to earlier versions, but sensitivity remained limited. True-positive cases had more strictly cortical lobar microbleeds than false-positive cases, strengthening the hypothesis that exact cortical versus juxtacortical localization contains additional etiologic information.

In a separate 54-person Mayo community and memory-clinic MRI-autopsy cohort without symptomatic intracerebral hemorrhage, probable CAA v2.0 sensitivity was 28.6%, specificity 65.3%, and AUC 0.47. Broadening to possible plus probable CAA raised sensitivity but sharply reduced specificity.

The correct synthesis is not that Boston v2.0 “failed.” It is that a classification rule optimized in one clinical spectrum should not be assumed to solve a different use case. Preclinical or cognitive-only CAA likely requires additional biomarkers and possibly different thresholds.

Interpreting individual features

Strictly lobar hemorrhagic lesions

They are among the strongest available MRI clues for CAA, especially when multiple and accompanied by cortical superficial siderosis or convexity subarachnoid hemorrhage. They remain probabilistic. Arteriolosclerosis, trauma, cavernous malformations, treatment effects, and other processes can create mimics or true lobar hemorrhagic lesions.

Cortical superficial siderosis

Its sulcal distribution and extent can be diagnostically and prognostically informative. It must be distinguished from other sources of chronic subarachnoid blood products and interpreted with susceptibility-sensitive sequences.

Centrum-semiovale perivascular spaces

Severe CSO-PVS add information in combination with a lobar hemorrhagic feature. Spatial pathology supports a relationship to overlying cortical CAA, but PVS also occur with aging and other small-vessel processes and do not directly measure fluid flow.

Multispot white-matter hyperintensity pattern

This is a qualifying feature within Boston v2.0, not a standalone CAA biomarker. Its incremental performance appears smaller than that of CSO-PVS in available community validation, and WMH remain etiologically nonspecific.

The Edinburgh criteria answer a narrower question

The Edinburgh CT and genetic model was developed for adults with lobar intracerebral hemorrhage. It combines subarachnoid extension, finger-like projections, and APOE e4. Among 62 lobar ICH cases in the autopsy cohort, the three-variable model achieved a c-statistic of 0.92; the proposed rule-in combination had 96% specificity. Its strengths are a masked pathology reference and practical CT morphology. Its boundaries are equally important: small development set, internal validation, genotype dependence, and no claim to classify non-ICH presentations.

A practical evidence-reading checklist

Before accepting any accuracy number, ask:

What was the intended context of use?
How were participants selected?
What clinical presentations were present?
Which MRI sequences and field strengths were required?
Was the index test rated blind to pathology?
Was the reference full autopsy, limited biopsy, hematoma evacuation, or clinical follow-up?
How long separated imaging and pathology?
Were mixed pathologies measured?
Were thresholds prespecified?
Was calibration reported, not only discrimination?
Has the result been reproduced outside the development ecosystem?

Current research implication

The next diagnostic advance is unlikely to come from adding one more nonspecific lesion to a checklist. Higher-value directions include exact lesion compartment, continuous spatial burden, vessel physiology, molecular/fluid information, acquisition-aware models, and multimodal validation against multiregion pathology. These models must be developed separately for symptomatic diagnosis, preclinical detection, treatment-risk estimation, and prognosis.