Sensitivity and Specificity of the Boston Criteria Version 2.0 for the Diagnosis of Cerebral Amyloid Angiopathy in a Community-Based Sample

Abstract

Background and Objectives

The Boston criteria are a set of clinical and neuroimaging features that enable accurate diagnosis of cerebral amyloid angiopathy (CAA) without invasive methods such as brain biopsies or autopsy. The last updates to the Boston criteria, named version 2.0, were recently released and incorporated new nonhemorrhagic MRI features. These criteria have been validated in symptomatic samples, with improved diagnostic yield. We set out to investigate the accuracy of the Boston criteria v2.0 for the diagnosis of CAA in a community-based sample.

Methods

Participants were recruited from longitudinal clinical-pathologic studies of aging conducted at the Rush Alzheimer's Disease Center in Chicago: the Religious Orders Study and the Rush Memory and Aging Project. Deceased participants with in vivo 3T MRI and detailed pathologic data available were included in the analysis. We compared the diagnostic yield of the current and earlier versions of the Boston criteria in our sample. Among those classified as probable CAA according to the Boston criteria v2.0, we investigated the ability of each neuroimaging marker to distinguish between false-positive and true-positive cases.

Results

In total, 134 individuals were included in the study (mean age = 82.4 ± 6.0 years; 69.4% F), and 49 of them were considered pathology-proven definite cases with CAA (mean age = 82.9 ± 6.0 years; 63.3% F). The Boston criteria versions 1.0 and 1.5 yielded similar sensitivity (26.5%, both), specificity (90.6% and 89.4%, respectively), and predictive values (negative: 68.1% and 67.9%; positive: 61.9% and 59.1%, respectively). The recently released Boston criteria v2.0 offered higher sensitivity (38.8%) and slightly lower specificity (83.5%). Among those classified as probable CAA (v2.0), pathology-proven true-positive cases had higher numbers of strictly cortical lobar microbleeds compared with false-positive cases (p = 0.004).

Discussion

Similar to findings from symptomatic samples, the inclusion of nonhemorrhagic neuroimaging markers in the updated Boston criteria offered a 12.3% gain in sensitivity among community-dwelling individuals, at the expense of a 5.9% drop in specificity. In cases with probable CAA, the cortical location of microbleeds may represent a promising distinguishing feature between true-positive and false-positive cases. Despite its improved performance, the diagnostic sensitivity of the updated criteria in a community-based sample remains limited.

Classification of Evidence

This study provides Class II evidence that the Boston criteria v2.0 accurately distinguishes people with CAA from those without CAA.

Introduction

Cerebral amyloid angiopathy (CAA) is a common neuropathologic finding characterized by the deposition of β-amyloid in the walls of cortical and leptomeningeal blood vessels.¹ CAA is the leading cause of recurrent lobar intracerebral hemorrhage (ICH) in older populations and has more recently been recognized as a highly prevalent vascular contributor to cognitive impairment and dementia.²

Initially developed in the 1990s, the Boston criteria are a set of clinical-radiologic features that allow for the diagnosis of CAA during life.^3,4 Per the initially proposed v1.0 criteria, the presence of multiple hemorrhages restricted to lobar regions in older individuals, not attributable to any other causes, is considered indicative of probable underlying CAA pathology. These criteria made CAA diagnosis no longer dependent on invasive brain biopsies or full postmortem examination and soon became the basis for clinical decision-making.⁵

The limited availability of confirmatory pathologic data derived from autopsies or biopsies is an obstacle to the validation of these criteria across different cohorts and clinical settings. As expected, the predictive value of the Boston criteria is dependent on how prevalent and advanced the disease is in the investigated cohort,⁵ performing better in symptomatic samples with ICH (sensitivity: 57.9%–76.9%; specificity: 87.5%–100%), in whom severe CAA pathology is more abundant.^5-8 By contrast, in symptomatic cohorts without ICH or in community-based cohorts, sensitivity (42.4% and 4.5%, respectively) and predictive values are lower.^5,9

In 2010, a modified version of the Boston criteria (v1.5) was released, incorporating cortical superficial siderosis (cSS), an emerging hemorrhagic MRI marker of CAA associated with a high risk of first-ever or recurrent ICH.⁷ Applied in a hospital-based cohort, these criteria showed improved sensitivity (from 57.9% to 71.1%),^5,7 without compromising specificity (95.5%).^5,7 Since then, extensive neuroimaging research led to the detection of new MRI findings associated with CAA. These advances prompted a multicenter effort to update, improve, and validate the Boston criteria (v2.0), taking advantage of 2 recently identified nonhemorrhagic MRI markers of CAA, namely severe perivascular spaces in the centrum semiovale (CSO-PVS) and the multispot subcortical pattern of white matter hyperintensities (WMH-MS).^10,11 Specifically, severe CSO-PVS refers to a high burden of MRI-visible PVS located in the centrum semiovale (>20 counted on 1 slice in 1 hemisphere). This finding has been found to be indicative of underlying CAA in ICH and memory-clinic patients.^12-14 The WMH-MS pattern is defined as the presence of ≥10 subcortical WMH circles or spots counted across the whole brain and was associated with lobar cerebral microbleeds (CMBs) and severe CSO-PVS in patients with ICH.¹⁵

Although these criteria have been tested among symptomatic patients from different centers,¹⁰ its performance in community-dwelling individuals remains to be explored. Assessing the performance of the Boston criteria v2.0 in population-based cohorts would, therefore, help better define its clinical generalizability and application in nonclinical settings. In addition, a more detailed evaluation of false-positive and false-negative cases may improve our understanding of how each neuroimaging feature contributes to pinpoint pathology-proven cases with CAA among community-dwelling older individuals.

The Religious Orders Study (ROS) and Rush Memory and Aging Project (MAP) are longitudinal ongoing clinical-pathologic studies that enroll older adults without dementia and have provided invaluable advances in the fields of neurodegenerative and vascular pathologies, offering unique opportunities to investigate pathologic correlates of incident neuroimaging findings.¹⁶ With that in mind, we sought to determine the diagnostic yield of the Boston criteria v2.0 in participants of the combined the Rush Religious Orders Study and Memory and Aging Project (ROSMAP) cohort. We hypothesized that the accuracy of these criteria would be lower in this community-based sample compared with reports from clinical-based samples. We further expanded on existing studies by taking a careful look into the false-positive and false-negative cases (i.e., individuals misdiagnosed with or without CAA through clinical-radiologic criteria) in an attempt to identify distinguishing features. As our primary research question, we aimed to investigate the accuracy of the Boston criteria v2.0 for the diagnosis of CAA in a community-based sample.

Methods

Study Cohort

Participants were selected from ROS and MAP studies, 2 ongoing longitudinal clinical-pathologic studies of aging conducted at the Rush Alzheimer's Disease Center of Rush University Medical Center in Chicago, Illinois.¹⁶ Both studies started in the 1990s, recruiting religious and lay individuals aged 65 years or older without dementia, who were submitted to annual clinical evaluations, and agreed to brain donation during death. Nuns, priests, brothers, and sisters across the United States are recruited into ROS,¹⁷ while MAP focuses on individuals from homes, retirement facilities, and subsidized housing across northeastern Illinois.¹⁸ Of note, these are not population-based studies, and individuals were recruited from specific sites.

Both ROS and MAP studies share a large common core of data, allowing for efficient merging of data.¹⁶ In vivo neuroimaging substudies started in 2009 (1.5T MRI) and 2012 (3T MRI), excluding individuals with cardiac pacemaker, certain metals in the body, and claustrophobia.

Within the ROSMAP cohort, we screened individuals recruited until March 2021, who had both in vivo 3T MRI and histopathologic data on CAA available for analysis. Exclusion criteria included motion artifacts and incomplete neuroimaging or pathologic data.

Standard Protocol Approvals, Registrations, and Patient Consents

The Religious Orders Study, Memory and Aging Project, and substudies have been approved by the institutional review board of Rush University Medical Center.¹⁶ Written informed consent and a signed anatomic gift act for brain donation was obtained from each participant upon enrollment.¹⁶

Clinical and Demographic Data

On enrollment, participants underwent a thorough and structured clinical evaluation performed by trained staff and a neuropsychological test technician. Demographic information, including age, sex, race, ethnicity, and education level were collected at baseline.¹⁶ Data on cognitive status and self-reported medical history of vascular risk factors were extracted from all study visits. Alzheimer dementia was diagnosed through the NINCDS-ADRDA criteria.¹⁹ Participants were considered cognitively normal if their neuropsychological test results were within the normal range and there was no evidence of dementia. Final cognitive diagnosis proximate to death was determined by a clinician with expertise in dementia using all years of clinical data and blinded to neuropathologic diagnosis.¹⁶

Neuroimaging Data

Examinations were performed for research purposes on 3T MRI scanners (Siemens Magnetom Trio or Philips Achieva) and included the following sequences: 3D T1-weighted imaging (slice thickness, 1 mm; in-plane resolution 1 × 1 mm), T2-weighted imaging (slice thickness, 2–3 mm; in-plane resolution 1.3–2 × 0.9–2 mm), fluid-attenuated inversion recovery (FLAIR) imaging (slice thickness, 4 mm; in-plane resolution 0.9 × 0.9–1.1 mm), and susceptibility-weighted imaging (SWI, slice thickness, 1–1.3 mm; in-plane resolution 0.7–1 × 0.7–1 mm). The scan protocols are summarized in eTable 1, links.lww.com/WNL/D230.

MRI Ratings

Conventional MRI markers of cerebral small vessel disease (SVD) were rated by a neuroradiologist with 7 years of experience (M.C.Z.Z.; observer 1), blinded to all clinical data, following the Standards for Reporting Vascular Changes on Neuroimaging recommendations.²⁰ The number and the location of CMBs,²¹ ICH, and the presence and extent of cSS were rated on SWI (Figure 1, A–C).⁷ CSO-PVS were assessed on T2-weighted images, using a previously validated scale,²² and severe CSO-PVS was defined as >20 visible PVS in the centrum semiovale of 1 hemisphere on 1 slice (Figure 1D).¹⁰ WMH-MS pattern was rated on FLAIR images, following previously published instructions.¹⁵ In brief, small round or ovoid subcortical T2-FLAIR hyperintense lesions were counted across the whole brain, and the presence of ≥10 lesions was indicative of WMH-MS pattern (Figure 1E).¹⁵ Specifically for those cases considered as probable CAA according to the Boston criteria v2.0, each lobar CMB was further rated as cortical, if located strictly within the cortical ribbon, or noncortical, if located in the subcortical or juxtacortical white matter (including the U-fibers).²¹

Figure 1. Current Neuroimaging Markers Used in the Boston Criteria v2.0 That Were Present in the ROSMAP Cohort.

Created with Biorender. PVS = perivascular spaces; WMH = white matter hyperintensities; ROSMAP = the Rush Religious Orders Study and Memory and Aging Project.

For interrater reliability analyses, images from a third of the participants (n = 31 randomly selected individuals) were also independently rated by a computer scientist with 5 years of experience in neuroimaging (NM). Original ratings performed by observer 1 were used for further analyses.

Boston criteria for Diagnosis of CAA

Based on the clinical and neuroimaging data extracted, participants were classified as cases with probable CAA and non-CAA cases, following the Boston criteria v1.0, v1.5, and v2.0 (Figure 1; eTable 2, links.lww.com/WNL/D230).

Of importance, the objective of our study was to assess the accuracy of the Boston criteria v2.0 in community-dwelling individuals, who are less likely to present with clinical features related to CAA compared with hospital-based cohorts. Therefore, we investigated the performance of the updated neuroimaging criteria in this population, regardless of the new clinical features other than age that have been added to the Boston criteria v2.0 (i.e. presentation with spontaneous ICH, transient focal neurologic episodes [TFNE], convexity subarachnoid hemorrhage [cSAH], or cognitive impairment/dementia).

Pathologic Data

The median (interquartile range [IQR]) postmortem interval (time between death and autopsy) was 7.9 (6.2–10.5) hours. Details on the autopsy procedure and neuropathologic analyses have been described elsewhere^23,24 and are available in eAppendix 1, links.lww.com/WNL/D230. In summary, CAA pathologic assessment was performed blinded to clinical and neuroimaging data and followed a modified protocol proposed by Love et al.^25,26 to create a continuous CAA-pathology variable, which was finally stratified into a 4-level semiquantitative scale: none, mild, moderate, and severe.

Following similar pathologic analyses,²⁷ individuals with moderate or severe CAA on pathology (according to the aforementioned semiquantitative scale) were considered cases with “definite CAA”, whereas individuals with none or mild CAA on pathology were considered “non-CAA” cases.

Statistical Analysis

To run between-group comparisons, individuals were stratified into definite CAA and non-CAA groups, as defined by pathology. The distribution of continuous variables was tested for normality using the Shapiro-Wilk test. Demographic, clinical, and neuroimaging findings were compared between the groups, using the χ2 or Fisher exact test and Mann-Whitney U test or t test, as appropriate. For categorical neuroimaging variables, we assessed classification measures (sensitivity and specificity) and univariable logistic regression (ORs with 95% CI) for a definite CAA diagnosis based on neuropathology. For continuous or ordinal variables, sensitivity and specificity were derived from receiver operating characteristic (ROC) analyses. We further calculated sensitivity, specificity, positive predictive values (PPV), and negative predictive values (NPV) for each of the 3 versions of the Boston criteria.

Similar analyses were run for comparing cases correctly and incorrectly classified as defined by the Boston criteria v2.0 in reference to pathologic data. For categorical neuroimaging variables (moderate/severe CSO-PVS, presence of cSS, and presence of WMH-MS pattern), interobserver agreement was estimated using Cohen kappa. Intraclass correlation coefficient (ICC) was used for ordinal variables (CMBs, stratified into 3 levels: 0, 1, and ≥2 CMBs).

The statistical significance level was set at 0.05 for all analyses. We used the Statistical Package for the Social Sciences (SPSS) version 20.0 (for IOS; SPSS Inc, Chicago, IL) and R (v3.5.3) to run the analyses.

Data Availability

Anonymized data used in this study can be made available for qualified investigators at reasonable request.

Results

Overall, of the 3,678 individuals enrolled in the combined ROSMAP cohort by March 2021, 2,068 were deceased, and 1,703 had pathologic data on CAA available (derived exclusively from autopsy samples). Among those, 160 had in vivo MRI scans, and a total of 134 were included in this study (mean age = 82.4 ± 6.0 years; 69.4% F). Fourteen individuals were excluded because of incomplete clinical/pathologic data, 10 because of incomplete neuroimaging data, and 2 because of motion artifacts (Figure 2). Within the final sample of 134 participants, 49 (36.6%) had moderate-to-severe CAA and 85 (63.4%) had none or mild CAA on pathology. In terms of clinical and pathologic data, we compared the 134 individuals included in our study against the other 1,569 deceased participants from the ROSMAP cohort, who had pathologic CAA data but were not included in the analyses (eTable 3, links.lww.com/WNL/D230). The participants included in our study were slightly older (age at baseline, median [IQR]: 82.8 years [78.7–86.2]) than those not included (80.7 years [75.3–85.2]), had lower prevalence of stroke (19/134 [14.3%] vs 380/1,569 [24.3%]), and more commonly used platelet inhibitor medications (113/134 [84.3%] vs 1,183/1,569 [75.4%]). In terms of cognitive performance, the included sample was enriched with participants with normal cognition compared with the nonincluded group (56/134 [41.8%] vs 488/1,569 [31.1%]). Note that we did not observe any significant differences between the groups in terms of the prevalence of moderate-to-severe CAA pathology (49/134 [36.6%] vs 572/1,569 [36.5%]) (eTable 3, links.lww.com/WNL/D230).

Figure 2. Flowchart Depicting the Selection of Participants Included in the Study.

Prevalence of Neuroimaging Markers of CAA

We observed a high prevalence of CAA-related MRI markers in our community-based sample. Specifically, 87 (64.9%) participants had at least 1 CMB, and most of them (74/87; 85.1%) had at least 1 lobar CMB. cSS was identified in 8 individuals (6%), while WMH-MS and severe CSO-PVS were identified in 49 (36.6%) and 59 (44%) participants, respectively (Table 1).

Table 1.

Comparison of Demographic, Clinical, and Neuroimaging Data Between Individuals With Definite CAA (Moderate-to-Severe CAA on Pathology) and Non-CAA Individuals (None-to-Mild CAA on Pathology)

	Total, n = 134	Non-CAA, n = 85	Definite CAA, n = 49	p Value
Demographics
Age at baseline, mean ± SD	82.4 ± 6.0	82.1 ± 6.1	82.9 ± 6.0	0.441
Age at MRI study, median [IQR]	88.4 [84.8–92.4]	88.5 [84.0–92.5]	88.2 [86.2–92.6]	0.447
Age at death, median [IQR]	91.8 [88.5–95.2]	91.6 [88.1–95.1]	91.8 [89.6–95.5]	0.453
Time MRI-autopsy, y, median [IQR]	2.9 [1.7–4.3]	2.8 [1.6–4.1]	3.1 [1.7–4.3]	0.659
Female sex, n (%)	93 (69.4)	62 (72.9)	31 (63.3)	0.242
Race				1.000
White	131 (97.8)	83 (97.6)	48 (98.0)
Black	2 (1.5)	1 (1.2)	1 (2)
Asian	1 (0.7)	1 (1.2)	0 (0)
Hispanic ethnicity	3 (2.2)	0 (0)	3 (6.1)	0.047a
Education, y, median [IQR]	16 [14–18]	16 [13.5–18]	16 [14–18]	0.378
Postmortem interval, h, median [IQR]	7.9 [6.2–10.5]	7.3 [5.9–9.5]	9.1 [6.5–13.1]	0.008a
Clinical
Hypertension, n (%)	94 (70.1)	59 (69.4)	35 (71.4)	0.806
Diabetes, n (%)	29 (21.6)	21 (24.7)	8 (16.3)	0.257
History of head injury with loss of consciousness, n (%)b	14 (10.9)	8 (9.8)	6 (12.8)	0.597
History of stroke, n (%)d	19 (14.3)	12 (14.3)	7 (14.3)	1.000
Anticoagulant, n (%)	42 (31.3)	28 (32.9)	14 (28.6)	0.599
Platelet inhibitor medication, n (%)	113 (84.3)	70 (82.4)	43 (87.8)	0.407
Cognitive status during death, n (%)				0.386
Normal cognition	56 (41.8)	32 (37.6)	24 (49.0)
MCI	25 (18.7)	18 (21.2)	7 (14.3)
AD	53 (39.6)	35 (41.2)	18 (36.7)
≥1ApoE4 allelec	26 (21.1)	12 (15.4)	14 (31.1)	0.040a
Neuroimaging
Presence of any ICH, n (%)	9 (6.7)	6 (7.1)	3 (6.1)	1.000
Presence of any CMB, n (%)	87 (64.9)	50 (58.8)	37 (75.5)	0.051
Lobar CMB presence, n (%)	74 (55.2)	40 (47.1)	34 (69.4)	0.012a
Deep CMB presence, n (%)	42 (31.3)	28 (32.9)	14 (28.6)	0.599
Cerebellar CMB presence, n (%)	27 (20.1)	20 (23.5)	7 (14.3)	0.199
Strictly lobar CMBs, n (%)	41 (30.6)	20 (23.5)	21 (42.9)	0.019a
Strictly deep CMBs, n (%)	7 (5.2)	6 (7.1)	1 (2)	0.422
Mixed pattern of CMBs, n (%)	35 (26.1)	22 (25.9)	13 (26.5)	0.934
Strictly cerebellar CMBs, n (%)	4 (3)	2 (2.4)	2 (4.1)	0.623
Total CMB count, median [IQR]	1 [0–4]	1 [0–4]	2 [0.5–4.5]	0.108
Lobar CMB count, median [IQR]	1 [0–2]	0 [0–2]	1 [0–3]	0.017a
≥2 strictly lobar CMB, n (%)	21 (15.7)	8 (9.4)	13 (26.5)	0.009a
Deep CMB count, median [IQR]	0 [0–1]	0 [0–1]	0 [0–1]	0.492
Cerebellar CMB count, median [IQR]	0 [0–0]	0 [0–0]	0 [0–0]	0.158
Presence of WMH-MS pattern, n (%)	49 (36.6)	29 (34.1)	20 (40.8)	0.438
Presence of severe CSO-PVS, n (%)	59 (44)	36 (42.4)	23 (46.9)	0.607
Presence of cSS, n (%)	8 (6)	4 (4.7)	4 (8.2)	0.463
Boston criteria
Probable CAA, Boston criteria v1.0, n (%)	21 (15.7)	8 (9.4)	13 (26.5)	0.009a
Probable CAA, Boston criteria v1.5, n (%)	22 (16.4)	9 (10.6)	13 (26.5)	0.016a
Probable CAA, Boston criteria v2.0, n (%)	33 (24.6)	14 (16.5)	19 (38.8)	0.004a

Abbreviations: AD = Alzheimer disease; ApoE = apolipoprotein E; CAA = cerebral amyloid angiopathy; CMB = cerebral microbleed; CSO-PVS = perivascular spaces in the centrum semiovale; cSS = cortical superficial siderosis; ICH = intracerebral hemorrhage; IQR = interquartile range; MCI = mild cognitive impairment; n = number; WMH-MS = white matter hyperintensities in a multispot subcortical pattern.

^aStatistically significant.

^b5 missing cases.

^c11 missing cases.

^d1 missing case.

Interobserver Agreement of MRI Ratings

Cohen kappa measurements of agreement for the presence of CMB (K = 1.0, p < 0.001), cSS (K = 0.674, p < 0.001), and WMH-MS pattern (K = 0.728, p < 0.001) were substantial to excellent. ICC for CMB count was substantial (ICC = 0.753; CI 95% 0.548, 0.872). Agreement was modest for the presence of moderate/severe CSO-PVS (K = 0.415, p = 0.018).

Between-Group Comparison

First, we compared clinical and neuroimaging findings between cases who met pathologic criteria for definite CAA (n = 49) and non-CAA cases (n = 85) (Table 1). The groups were similar in terms of age and sex. The definite-CAA group had a slightly higher prevalence of Hispanic ethnicity than non-CAA cases (3/49 [6.1%] vs 0/85 [0%]; p = 0.047). Cases with definite CAA also had higher postmortem intervals (hours [IQR]; 9.1 [6.5–13.1]; p = 0.008) compared with non-CAA cases (7.3 [5.9–9.5]). Although a longer postmortem interval has been associated with worse tissue quality, it does not significantly affect the detection of β-amyloid with immunohistochemistry and hence likely had little, if any, effect on our results. In terms of neuroimaging, lobar CMBs were, as expected, more frequent (p = 0.012) and found in higher numbers (p = 0.017) among cases with definite CAA (Table 1). In univariable analysis, among several MRI markers, only the presence of lobar CMBs (OR [95% CI] = 2.550 [1.214–5.355], p = 0.013) and the presence of ≥2 strictly lobar CMBs (OR [95% CI] = 3.476 [1.323–9.128], p = 0.011) increased the risk of CAA pathology. In ROC analyses, only the number of lobar CMBs distinguished cases with definite CAA from non-CAA cases (area under the curve [AUC] [95% CI] = 0.618 [0.520–0.716], p = 0.023).

Diagnostic Yield of Different Versions of the Boston Criteria

Next, we evaluated the performance of the Boston criteria in the combined ROSMAP cohort. The 3 versions of the Boston criteria showed slightly different diagnostic yields. Versions 1.0 and 1.5 yielded similar sensitivity (both 26.5%), specificity (90.6% and 89.4%, respectively), predictive values (NPV: 68.1% and 67.9%, respectively; PPV: 61.9% and 59.1%, respectively), and AUCs (AUC [95% CI] = 0.586 [0.482–0.689], p = 0.100; 0.580 [0.476–0.683], p = 0.125, respectively) (Table 2, Figure 3). The inclusion of the new neuroimaging criteria of version 2.0 led to a 12.3% increase in sensitivity (38.8%), at the expense of a 5.9% drop in specificity (83.5%), with an NPV of 70.3%, a PPV of 57.6%, and a higher AUC [95%] (0.612 [0.509–0.714], p = 0.032) (Table 2, Figure 3). Twelve cases previously considered as possible CAA in v1.5 were upgraded to cases with probable CAA in v2.0, 6 of which were true-positive cases. Of note, none of the upgraded cases with probable CAA in v2.0 had multifocal cSAH/cSS, and all of them had 1 lobar hemorrhagic feature in addition to at least 1 white matter finding. Of the white matter features, the addition of CSO-PVS raised the sensitivity to 38.8%, whereas the addition of WMH-MS raised it to 34.7%, while specificity reached 84.7% and 85.9%, respectively.

Table 2.

Diagnosis of CAA in a Community-Based Sample, Using Different Versions of the Boston Criteria

Clinical-radiologic criteria	Boston criteria v1.0			Boston criteria v1.5			Boston criteria v2.0
	Pathologic criteria			Pathologic criteria			Pathologic criteria
	n	Definite CAA	Non-CAA	n	Definite CAA	Non-CAA	n	Definite CAA	Non-CAA
Probable CAA	21	13	8	22	13	9	33	19	14
Possible CAA	20	8	12	20	9	11	36	9	27
Non-CAA	93	28	65	92	27	65	65	21	44
Total	134	49	85	134	49	85	134	49	85

Diagnostic yield	Probable CAA*	Prob. + Poss. CAA†	Probable CAA*	Prob. + Poss. CAA†	Probable CAA*	Prob. + Poss. CAA†
Sensitivity (95% CI)	26.5% (14.2, 38.9)	42.9% (29.0, 56.7)	26.5% (14.2, 38.9)	44.9% (31.0, 58.8)	38.8% (25.1, 52.4)	57.1% (43.3, 71.0)
Specificity (CI 95%)	90.6% (84.4, 96.8)	76.5% (67.5, 85.5)	89.4% (82.9, 96)	76.5% (67.5, 85.5)	83.5% (75.6, 91.4)	51.8% (41.1, 62.4)
Positive predictive value (CI 95%)	61.9% (41.1, 82.7)	51.2% (35.9, 66.5)	59.1% (38.6, 79.6)	52.4% (37.3, 67.5)	57.6% (40.7, 74.4)	40.6% (29.0, 52.2)
Negative predictive value (CI 95%)	68.1% (59.6, 76.7)	69.9% (60.6, 79.2)	67.9% (59.2, 76.5)	70.7% (61.4, 80.0)	70.3% (61.4, 79.2)	67.7% (56.3, 79.1)
AUC (CI 95%)	0.586 (0.482, 0.689)	0.597 (0.495, 0.698)	0.580 (0.476, 0.683)	0.607 (0.505, 0.708)	0.612 (0.509, 0.714)	0.545 (0.443, 0.646)

Abbreviations: AUC = area under the curve; CAA = cerebral amyloid angiopathy.

* In the columns labeled “Probable CAA” we display the diagnostic yield of probable CAA compared to the rest of the sample, which included non-CAA and possible CAA cases (according to the Boston criteria).

^† In the columns labeled “Probable + Possible CAA” we display the diagnostic yield of probable + possible CAA compared to the rest of the sample, which included non-CAA cases (according to the Boston criteria).

Figure 3. Receiver Operating Characteristic Curves Comparing the Diagnostic Performance of the 3 Versions of the Boston Criteria.

We ran a sensitivity analysis further excluding the 14 individuals with a history of head injury with loss of consciousness (n = 120). Results remained similar for Boston criteria versions 1.0 (sensitivity 27.5%; specificity 92.2%; PPV 66.7%; NPV 69.6%; and AUC [95% CI] = 0.601 [0.491–0.710], p = 0.068), 1.5 (sensitivity 27.9%; specificity 90.9%; PPV 63.2%; NPV 69.3%; and AUC [95% CI] = 0.594 [0.484–0.704], p = 0.088), and 2.0 (sensitivity 39.5%; specificity 87%; PPV 63%; NPV 72%; and AUC [95% CI] = 0.633 [0.525–0.741], p = 0.016).

Analysis of False-Positive Cases

Among those classified as probable CAA according to the Boston criteria v2.0, the true-positive cases had higher numbers of lobar CMBs that were rated as cortical (p = 0.004) (Table 3). The ability of each neuroimaging marker to distinguish between false and true positives was further assessed with ROC analysis. The number of cortical lobar CMBs (AUC [95% CI] = 0.789 [0.622–0.957], p = 0.005) and the number of noncortical lobar CMBs (AUC [95% CI] = 0.192 [0.047–0.337], p = 0.003) showed high discriminatory power within the subset of cases meeting criteria for CAA during life (Figure 4, eTable 4, links.lww.com/WNL/D230).

Table 3.

Detailed Demographic, Clinical, and Neuroimaging Description of True-Positive and False-Positive Cases, According to the Boston Criteria v2.0

	False positives	True positives	p Value
	n = 14	n = 19
Demographics
Age at baseline, mean ± SD	81.1 ± 3.5	81.6 ± 4.2	0.753
Age at MRI study, mean ± SD	89.1 ± 4.6	88.8 ± 4.9	0.864
Age at death, mean ± SD	91.8 ± 4.3	91.9 ± 4.6	0.921
Time MRI-autopsy, y, mean ± SD	2.6 ± 1.6	3.1 ± 1.7	0.454
Female sex, n (%)	10 (71.4)	12 (63.2)	0.719
Race			1.000
White	14 (100)	18 (94.7)
Black	0 (0)	1 (5.3)
Hispanic ethnicity	0 (0)	2 (10.5)	0.496
Education, y, median [IQR]	17 [12.8–20]	16 [14–18]	0.627
Postmortem interval, hours, median [IQR]	7.5 [6.0–8.9]	9 [6.5–20.8]	0.123
Clinical
Hypertension, n (%)	11 (78.6)	14 (73.7)	1.000
Diabetes, n (%)	3 (21.4)	3 (15.8)	1.000
History of head injury with loss of consciousness, n (%)	4 (28.6)	2 (10.5)	0.363
History of stroke, n (%)c	2 (15.4)	4 (21.1)	1.000
Anticoagulant medication, n (%)	5 (35.7)	7 (36.8)	0.947
Platelet inhibitor medication, n (%)	12 (85.7)	15 (78.9)	1.000
Cognitive status during death			0.142
Normal cognition	5 (35.7)	9 (47.4)
MCI	3 (21.4)	0 (0)
AD	6 (42.9)	10 (52.6)
≥1ApoE4 alleleb	3 (23.1)	8 (42.1)	0.450
Neuroimaging
Presence of any ICH, n (%)	0 (0)	1 (5.3)	1.000
Cerebellar CMB presence, n (%)	2 (14.3)	2 (10.5)	1.000
Total CMB count, median [IQR]	2 [1–3]	2 [1–5]	0.483
Lobar CMB count, median [IQR]	2 [1–3]	2 [1–4]	0.506
Lobar cortical CMB count, median [IQR]	0 [0–0]	1 [0–4]	0.004a
Lobar noncortical CMB count, median [IQR]	2 [1–2]	0 [0–1]	0.002a
≥2 strictly lobar CMB, n (%)	8 (57,1)	13 (68.4)	0.506
Cerebellar CMB count, median [IQR]	0 [0–0]	0 [0–0]	0.843
Presence of WMH-MS pattern, n (%)	6 (42.9)	8 (42.1)	0.966
Presence of severe CSO-PVS, n (%)	7 (50)	10 (52.6)	0.881
Presence of cSS, n (%)	1 (7.1)	2 (10.5)	1.000

Abbreviations: AD = Alzheimer disease; ApoE = apolipoprotein E; CAA = cerebral amyloid angiopathy; CMB = cerebral microbleed; CSO-PVS = perivascular spaces in the centrum semiovale; cSS = cortical superficial siderosis; ICH = intracerebral hemorrhage; IQR = interquartile range; MCI = mild cognitive impairment; n = number; WMH-MS = white matter hyperintensities in a multispot subcortical pattern.

^aStatistically significant.

^b1 missing case.

^c1 missing case.

Figure 4. Schematic and Example Images of Cortical and Noncortical CMB Locations.

Images named as a represent the SWI. We further coregistered the 3D T1-weighted images to the SWI, overlapped both images, and then reduced the opacity of the SWI so that the cortical boundaries well-appreciated on T1 could be seen through and depicted in relation to the CMB (better appreciated on SWI). These merged images were named as b. Images A and B display lobar CMBs located strictly within the cortical ribbon, whereas images C and D display lobar CMBs located outside the cortex. CMBs were more commonly found within the cortical ribbon in true-positive cases, whereas in false-positive cases, they were more commonly observed outside the cortical ribbon, in the subcortical and juxtacortical white matter. Created with Biorender.

Analysis of False-Negatives Cases

Among those classified as nonprobable CAA according to the Boston criteria v2.0, demographic, clinical, and neuroimaging features were overall similar. The false-negative cases had higher postmortem intervals (p = 0.041) (eTable 5, links.lww.com/WNL/D230). While statistical tests did not reach significance, we observed numerically higher prevalence of WMH-MS pattern, severe CSO-PVS, and cSS and greater total and lobar CMB counts among false-negative cases (eTable 5, links.lww.com/WNL/D230).

This study provides Class II evidence that the Boston criteria v2.0 accurately distinguishes people with CAA from those without CAA.

Discussion

Several key findings emerge from this study applying the v2.0 updates of the Boston criteria among community-dwelling older individuals. The inclusion of white matter neuroimaging criteria offered a gain in sensitivity, at the expense of somewhat reduced specificity. Among cases with probable CAA, the cortical location of lobar CMBs may represent a promising distinguishing feature between true-positive and false-positive cases. Despite its improved performance, the diagnostic accuracy of the updated criteria in a community-based sample remains overall moderate.

In the ROSMAP cohort, the inclusion of white matter nonhemorrhagic markers in the Boston criteria led to an increased detection of cases with true definite CAA, accompanied by an increase in false positives. Specifically, 38.8% of individuals harboring moderate-to-severe CAA on pathology fulfilled criteria for probable CAA. This represents a 12.3% absolute increase in sensitivity compared with the Boston criteria v1.5. The trade-off was a small reduction in specificity. The updated criteria accurately identified 83.5% of individuals without moderate-to-severe CAA on pathology, 5.9% less than the v1.5 criteria, but fairly similar to values reported in symptomatic samples (81.5%–89.5%).¹¹ These changes in sensitivity (+12.3%) and specificity (−5.9%) with overall increase in accuracy as measured by AUC compared with the earlier version of the Boston criteria (v1.5) are also similar to findings from symptomatic samples (+4.6% to +12.2% and −1.9% to −10.5%, respectively).¹¹ Of importance, in our cohort, these changes were exclusively due to the inclusion of white matter markers because, as expected for a community-based sample, multifocal cSS was an extremely rare feature. Our observation that CSO-PVS added more sensitivity in comparison with WMH-MS pattern is also in line with previous studies, which indicate a fairly marginal added performance from WMH-MS pattern.¹¹

Of importance, all versions of the Boston criteria yielded substantially lower sensitivity levels in our sample compared with symptomatic cohorts. For instance, particularly high sensitivity values were reported among those presenting with ICH (v2.0 90.2%).¹¹ Variability in the accuracy of diagnostic tests across different cohorts is a well-established phenomenon, linked to the prevalence and severity of the disease and its biomarkers within each sample. For instance, strictly lobar CMBs and the Boston criteria v1.0 offered higher diagnostic accuracy for CAA in hospital-based cohorts compared with that for population-based cohorts.⁹ The Boston criteria v1.0 was 88% specific for the diagnosis of CAA in a sample of 47 individuals from the Framingham study (a population-based cohort)⁹, similar to our specificity rate of 90.6%. Although they reported a sensitivity of only 4.5%, the same criteria detected up to 26.5% of individuals with pathology-proven CAA in our sample. This difference is likely due to better quality of neuroimaging data (3T SWI vs 1.5T T2* gradient recalled echo), through which we could have detected more CMBs.

Of importance, false positives contribute to lower the specificity levels of the Boston criteria among community-based cohorts, possibly driven by neuroimaging findings related to other co-occurring cerebral SVDs and/or traumatic events. Among several neuroimaging features, we observed that the cortical location of CMBs represents a promising feature that may distinguish between true-positive and false-positive cases. Considering the pathophysiology of CAA, which involves the deposition of β-amyloid in the walls of strictly cortical and leptomeningeal blood vessels, it is expected that most CAA-related CMBs will occur in the cortical ribbon, rather than subcortical or juxtacortical white matter. This hypothesis is further supported by several in vivo and ex vivo MRI studies^28-30 that indicate the cortex as the main site of CAA-related bleeding. In a study that scanned 5 individuals with CAA in a 7T MRI scanner, 169 of the total 170 CMBs visualized were located in the cortex.³⁰ Although a similar distinction was possible in our high-resolution 3D SWI sequences, it may be particularly challenging in clinical scans. The feasibility and reliability of distinguishing between cortical and noncortical lobar CMBs in routine clinical protocols has not been addressed yet and remains unclear. Prior Boston criteria validation studies have mainly used clinical imaging data and have thus not made this distinction. While our results regarding CMB location are still preliminary, they point to a promising MRI feature that might help increase the specificity of clinical-radiologic criteria for CAA in the future.

Our results differed from population-based studies in terms of the prevalence of neuroimaging markers of CAA. While the prevalence of CMBs and cSS ranged 3.1%–38% and 0.4%–0.9% in previous studies,³¹ we detected the same markers in 64.9% and 6% of our population, respectively. Several aspects may explain this discrepancy. Our sample included highly aged, predominantly female individuals, with a mean age of 88.5 years at the moment of the MRI scan, most of whom had some degree of cognitive impairment and who, in general, displayed other concomitant cerebral SVD markers. Because some degree of CAA pathology is common in older populations, by including more aged participants, we could be selecting a sample more prone to display imaging findings of CAA. Higher CMB rates have been consistently observed in older samples and in cohorts with other concomitant cerebral SVD markers at play.^31,32 Among patients with subcortical vascular dementia, CMBs were detected in 84.9% of cases using T2* gradient recalled echo images from a 1.5T scanner.³³ Moreover, our participants were all imaged on 3T scanners using highly sensitive sequences (3D SWI) read by an experienced neuroradiologist. It is well established that detection rates of CMBs increase with MRI field strength and are higher in SWI compared with T2* gradient recalled echo.^32,34 Furthermore, our sample was also enriched with clinical features that increase the risk of brain hemorrhages, such as hypertension (70.1%), history of head injury (10.9%), and use of anticoagulant (31.3%) and antiplatelet (84.3%) medications, which may also explain the high prevalence of hemorrhagic markers observed.

Data on PVS are less abundant, and different visual scores available pose an obstacle to direct comparisons.³⁵ Nonetheless, using similar scales and sequences, severe CSO-PVS has been detected in 22.6% of individuals from a population-based study,³⁶ in 21.7% of cognitively normal participants from the Alzheimer's Disease Neuroimaging Initiative database,³⁷ and in 43.8% of patients with symptomatic lobar ICH.¹⁵ The strong association between CSO-PVS and age and with lobar CMBs might also explain our higher rates of severe CSO-PVS (44%).³⁶ No studies to date have assessed the prevalence of WMH-MS pattern among community-dwelling individuals, but in patients with symptomatic ICH, the prevalence ranged between 16.8% and 29.8%.¹⁵ The overall high prevalence of MRI markers in our sample may also indicate that other competing cerebral SVDs may be at play, contributing to lower diagnostic accuracy.

The ability to diagnose CAA in community-based individuals, including those who have no clinical symptoms, is a research priority. Clinical symptoms of CAA are believed to occur later in the disease course.³⁸ Identifying presymptomatic individuals might allow interventions to be started at earlier disease stages when the progression of CAA might be most preventable. Of importance, the v2.0 criteria were derived and validated exclusively in symptomatic individuals presenting with either ICH, cognitive impairment, or TFNEs.¹¹ For this reason, symptomatic presentation was considered a requirement for the diagnosis. This analysis suggests that the neuroimaging features of the Boston criteria v2.0 might be extended to community-based individuals with reasonable diagnostic accuracy. Our findings suggest that the Boston criteria provide information on the likelihood of CAA in older individuals without stroke-like symptoms and are suitable for research purposes. However, given the limited diagnostic yield, applying these criteria in routine clinical practice to identify CAA in asymptomatic individuals is likely not warranted.

Our study has limitations. Although the inclusion criteria requiring availability of postmortem tissue and in vivo MRI may have caused selection bias toward participants more likely to harbor neuropathology, we observed no significant difference in the distribution of CAA pathology features between included and nonincluded individuals. Our relatively small sample size and the high prevalence of neuroimaging markers of CAA might reduce the generalizability of our findings across other, less affected, community-based cohorts. Furthermore, although our results do improve understanding on the role of the Boston criteria in diagnosing CAA among community-based individuals, the highly aged nature of our sample hampers the translation of such results to clinical practice, when early diagnosis might be particularly crucial. The increased prevalence of other possible underlying causes of neuroimaging abnormalities among older individuals might also have affected the diagnostic performance of the criteria in our sample. Our pathologic data were solely derived from postmortem tissue, analyzed using a semiquantitative pathologic scale²⁶ based on a consensus protocol for the assessment of CAA in autopsy tissue.²⁵ Prior validation studies for the Boston criteria have used both biopsy/hematoma evacuation samples and full-brain autopsies, relying predominantly on the modified Vonsattel rating system to assess CAA severity.³⁹ In samples with autopsy-derived pathologic data rated using the Vonsattel scale,¹¹ the Boston criteria v2.0 also showed improved sensitivity with similar specificity compared with the Boston criteria v1.5. Still, studies comparing the 2 pathologic rating systems^25,39 for equivalence in terms of diagnostic accuracy are largely missing. Nonetheless, several studies, including the recent validation of diagnostic criteria for CAA based on CT and genetic findings,²⁷ have used a pathologic consensus similar to the one we used in our analyses.²⁵ Although some studies support MRI-derived WMH-MS pattern as a strong predictor of CAA pathology, this particular neuroimaging marker has been seldomly investigated in the literature and has yet to be tested and validated by external researchers.

Finally, our findings indicate that the diagnostic accuracy of the Boston clinical-radiologic criteria in community-based cohorts improved with the recent v2.0 updates, but sensitivity remains limited. Most asymptomatic individuals with moderate-to-severe CAA on pathology remain undetectable through noninvasive techniques, warranting the development and validation of other biomarkers and criteria capable of detecting early changes in the disease course that precede hemorrhagic and clinical manifestations. Future studies should also focus on investigating the feasibility, reliability, and accuracy of neuroimaging features that might decrease false-positive rates, such as the strictly cortical location of CMBs. Their feasibility and reliability in the context of routine clinical scans should also be investigated. Finally, the field could further benefit from studies assessing the role of other markers, for instance, derived from PET and CSF data, in improving the diagnostic yield of the Boston criteria among community-dwelling individuals.

Glossary

AUC

area under the curve

CAA

cerebral amyloid angiopathy

CMBs

cerebral microbleeds

cSAH

convexity subarachnoid hemorrhage

CSO-PVS

perivascular spaces in the centrum semiovale

cSS

cortical superficial siderosis

FLAIR

fluid-attenuated inversion recovery

ICC

intraclass correlation coefficient

ICH

intracerebral hemorrhage

IQR

interquartile range

MAP

Memory and Aging Project

NPV

negative predictive values

PPV

positive predictive values

ROC

receiver operating characteristic

ROS

Religious Orders Study

ROSMAP

the Rush Religious Orders Study and Memory and Aging Project

SPSS

Statistical Package for the Social Sciences

SVD

small vessel disease

SWI

susceptibility-weighted imaging

TFNE

transient focal neurologic episodes

WMH-MS

multispot subcortical pattern of white matter hyperintensities

Appendix. Authors

Name	Location	Contribution
Maria Clara Zanon Zotin, MD, PhD	J. Philip Kistler Stroke Research Center, Department of Neurology, Massachusetts General Hospital, Harvard Medical School, Boston, MA; Center for Imaging Sciences and Medical Physics, Department of Medical Imaging, Hematology and Clinical Oncology, Ribeirão Preto Medical School, University of São Paulo, Ribeirão Preto, SP, Brazil	Drafting/revision of the article for content, including medical writing for content; major role in the acquisition of data; study concept or design; and analysis or interpretation of data
Nazanin Makkinejad, PhD	J. Philip Kistler Stroke Research Center, Department of Neurology, Massachusetts General Hospital, Harvard Medical School, Boston, MA	Drafting/revision of the article for content, including medical writing for content; major role in the acquisition of data
Julie A. Schneider, MD	Rush Alzheimer's Disease Center, Rush University Medical Center, Chicago, IL	Drafting/revision of the article for content, including medical writing for content; major role in the acquisition of data
Konstantinos Arfanakis, PhD	Rush Alzheimer's Disease Center, Rush University Medical Center, Chicago, IL; Department of Biomedical Engineering, Illinois Institute of Technology, Chicago, IL	Drafting/revision of the article for content, including medical writing for content; major role in the acquisition of data
Andreas Charidimou, MD, PhD	J. Philip Kistler Stroke Research Center, Department of Neurology, Massachusetts General Hospital, Harvard Medical School, Boston, MA	Drafting/revision of the article for content, including medical writing for content; study concept or design; and analysis or interpretation of data
Steven M. Greenberg, MD, PhD	J. Philip Kistler Stroke Research Center, Department of Neurology, Massachusetts General Hospital, Harvard Medical School, Boston, MA	Drafting/revision of the article for content, including medical writing for content; study concept or design; and analysis or interpretation of data
Susanne J. van Veluw, PhD	J. Philip Kistler Stroke Research Center, Department of Neurology, Massachusetts General Hospital, Harvard Medical School, Boston, MA	Drafting/revision of the article for content, including medical writing for content; major role in the acquisition of data; study concept or design; and analysis or interpretation of data

Study Funding

This work was supported by the NIH/NIA (R00 AG059893 to S.J.v.V.) and the American Heart Association/Bugher Foundation (814,728 to S.J.v.V.), the NIH/NINDS (UF1NS100599 to K.A.), and the NIH/NIA (R01AG064233, R01AG052200, P30AG072975 to K.A.).

Disclosure

The authors report no relevant disclosures. Go to Neurology.org/N for full disclosures.