Identification of Distinct Brain MRI Phenotypes and Their Association With Long-Term Dementia Risk in Community-Dwelling Older Adults

Abstract

Background and Objectives

Individual brain MRI markers only show at best a modest association with long-term occurrence of dementia. Therefore, it is challenging to accurately identify individuals at increased risk for dementia. We aimed to identify different brain MRI phenotypes by hierarchical clustering analysis based on combined neurovascular and neurodegenerative brain MRI markers and to determine the long-term dementia risk within the brain MRI phenotype subgroups.

Methods

Hierarchical clustering analysis based on 32 combined neurovascular and neurodegenerative brain MRI markers in community-dwelling individuals of the Age-Gene/Environment Susceptibility Reykjavik Study was applied to identify brain MRI phenotypes. A Cox proportional hazards regression model was used to determine the long-term risk for dementia per subgroup.

Results

We included 3,056 participants and identified 15 subgroups with distinct brain MRI phenotypes. The phenotypes ranged from limited burden, mostly irregular white matter hyperintensity (WMH) shape and cerebral atrophy, mostly irregularly WMHs and microbleeds, mostly cortical infarcts and atrophy, mostly irregularly shaped WMH and cerebral atrophy to multiburden subgroups. Each subgroup showed different long-term risks for dementia (min–max range hazard ratios [HRs] 1.01–6.18; mean time to follow-up 9.9 ± 2.6 years); especially the brain MRI phenotype with mainly WMHs and atrophy showed a large increased risk (HR 6.18, 95% CI 3.37–11.32).

Discussion

Distinct brain MRI phenotypes can be identified in community-dwelling older adults. Our results indicate that distinct brain MRI phenotypes are related to varying long-term risks of developing dementia. Brain MRI phenotypes may in the future assist in an improved understanding of the structural correlates of dementia predisposition.

Introduction

Most older adults have brain changes on MRI, such as cerebral atrophy or manifestations of cerebral small vessel disease (SVD).¹ These brain MRI markers mostly represent late resulting damage of different underlying pathologies and are therefore largely unspecific. This makes differentiation of underlying pathology based on brain MRI challenging.¹ Moreover, some brain abnormalities detected on MRI are regarded as related to normal aging. It is currently unknown if specific brain MRI phenotypes represent an increased risk for dementia. The ability to determine an individual's risk for dementia based on MRI may in the future be useful to determine patient prognosis and may aid in patient selection for future treatment studies.

Common brain MRI markers of neurovascular and neurodegenerative diseases are white matter hyperintensities (WMHs), lacunes, microbleeds, enlarged perivascular spaces, and cerebral atrophy. These brain MRI markers have been studied previously and are associated with the occurrence of dementia.^2-6 However, individual brain MRI markers only show at best a modest association with long-term occurrence of dementia.⁴ It, therefore, remains challenging to identify individuals who are at increased risk to develop dementia.² Because of heterogenous etiology and mixed pathologies, methods combining different brain MRI markers into one model may likely aid in a more detailed characterization of, potential prognostically relevant, so-called brain MRI phenotypes.

In a previous study within our group (in a different cohort), we aimed to detect an increased stroke and mortality risk in patients with manifest arterial disease and analyzed brain MRI markers in a combined way using a hierarchical clustering approach, resulting in the identification of different brain MRI phenotypes.⁷ In that study, distinct brain MRI phenotypes were detected that were associated with a different risk of future stroke and mortality. These brain MRI phenotypes can aid to identify the structural correlates of predisposition to different disease outcomes. The association of distinct brain MRI phenotypes with long-term dementia risk remains unknown.

We therefore aimed to identify different brain MRI phenotypes in community-dwelling individuals by combined hierarchical clustering analysis based on neurovascular and neurodegenerative brain MRI markers. Within each of these brain MRI phenotype subgroups, we determined the long-term dementia risk.

Methods

Participants and Study Design

The data set used for the current analysis was acquired as part of the population-based Age-Gene/Environment Susceptibility (AGES) Reykjavik Study.⁸ The cohort study was originally established in 1967 to prospectively study cardiovascular disease in a random sample from the general population in Iceland. Participants were born between 1907 and 1935 and were living in Reykjavik in 1967. Remaining participants of the cohort were randomly selected for a follow-up and underwent a baseline brain MRI scan between 2002 and 2006.

Baseline diagnosis of dementia was assessed in a 3-step process, as described previously.⁹ In short, participants underwent the Mini-Mental State Examination and the Digit Symbol Substitution Test. Participants were administered a second battery of diagnostic tests based on positive results in the previous tests and possibly a third stage, which included neurologic tests and a proxy interview.⁹ Based on these tests, participants were considered to have normal cognition, mild cognitive impairment, or dementia at baseline. Dementia diagnosis based on the Diagnostic and Statistical Manual, Fourth Edition, guidelines was made in a consensus meeting with a geriatrician, neurologist, neuropsychologist, and neuroradiologist.

Education level and smoking status were collected using questionnaires. The highest completed education level (primary school, secondary school, college, and university) was entered. Participants who never smoked were categorized as nonsmokers, participants who smoked regularly and at least 100 cigarettes or 20 cigars in a lifetime were categorized as former smokers, and participants who currently smoke were categorized as current smokers. Height (in centimeters) and weight (in kilograms) were measured and used to calculate body mass index. Hypertension was based on self-report, use of antihypertensive medication, or based on the measurements of systolic blood pressure >140 mm Hg and/or diastolic blood pressure >90 mm Hg. A standard mercury sphygmomanometer was used to measure systolic and diastolic blood pressure; the mean of 2 measurements was calculated. Diabetes mellitus was based on self-report of diabetes, use of antidiabetic medication, or fasting blood glucose level >7.0 mmol/L. Coronary artery disease was based on self-report plus the use of nitrates or evidence of a myocardial infarction on electrocardiogram.

Participants were followed from the date of the baseline MRI scan until diagnosis of dementia, loss to follow-up, or end of follow-up. Loss to follow-up means that the participants died or could not be contacted. Tracking for dementia diagnosis was done through vital statistics and hospital records and by the nursing home and home-based resident assessment instrument.¹⁰ The dementia follow-up of the AGES Reykjavik Study was concluded in 2015 (end of follow-up). The inclusion and exclusion of participants from the AGES Reykjavik Study for the current study is illustrated in Figure 1. For example, participants who were demented at baseline were excluded.

Figure 1. Flowchart Illustrating the Inclusion and Exclusion of Participants.

WMH = white matter hyperintensity.

Standard Protocol Approvals, Registrations, and Patient Consents

The study was approved by the Icelandic National Bioethics Committee, VSN:00-063, and the institutional review board responsible for the National Institute on Aging research; all participants signed for informed consent.

MRI Scanning Protocol

A baseline brain MRI scan was acquired on a 1.5 T Signa TwinSpeed system (General Electric Medical Systems, Waukesha, WI). The MRI protocol included a fluid-attenuated inversion recovery (FLAIR) sequence (repetition time = 8,000 milliseconds; time to echo = 100 milliseconds; inversion time = 2,000 milliseconds; field of view = 220 mm; voxel size = 0.86 × 0.86 × 3.00 mm³; interleaved slices) and a T1-weighted sequence (repetition time = 21 milliseconds; time to echo = 8 milliseconds; field of view = 240 mm; slice thickness = 1.5 mm; voxel size = 0.94 × 0.94 × 1.50 mm³).¹¹ A T2*-weighted gradient echo-type echo planar sequence (time to echo = 50 milliseconds; repetition time = 3,050 milliseconds; flip angle = 90°; field of view = 220 mm; matrix = 256 × 256) and a proton density/T2-weighted fast-spin echo sequence (time to first echo = 22 milliseconds; time to second echo = 90 milliseconds; repetition time = 3,220 milliseconds; echo train length = 8; flip angle = 90°; field of view = 220 mm; matrix = 256 × 256) were also part of the MRI protocol.

Brain MRI Markers

The brain MRI markers included to determine the brain MRI phenotypes were brain tissue volumes for the estimation of brain atrophy, WMH volumes, WMH shape markers, brain infarcts, microbleeds, and enlarged perivascular spaces.

Gray matter, white matter, CSF, and WMH were segmented automatically with a modified algorithm based on the Montreal Neurological Institute pipeline.¹² Intracranial volume was calculated by adding gray matter, white matter, CSF, and WMH volumes.¹³ Brain parenchymal fraction, white matter fraction, gray matter fraction, and lateral ventricle fraction were calculated by expressing the volumes as a fraction of intracranial volume.

Volumes of periventricular/confluent, deep, and total WMH were determined automatically using an in-house developed pipeline.¹⁴ Moreover, volumes of deep and periventricular WMH per lobe were calculated using a mask to delineate the lobes. WMH shape markers (fractal dimension, solidity, convexity, concavity index, and eccentricity¹⁵) were calculated, as previously described.¹⁴ A description of the shape markers, as well as their corresponding formulas, can be found in eTable 1 and eFigure 1. Brain infarcts (subcortical, cerebellar, and cortical infarcts), microbleeds, and enlarged perivascular spaces were visually scored.⁹ Microbleeds were first scored by neuroradiologists and then by trained radiographers.¹¹ Infarcts were defined as parenchymal defects with a signal intensity that is isointense to that of CSF on all MRI sequences (i.e., FLAIR, T2-weighted, proton density-weighted).⁹ Cortical infarcts were defined as infarcts involving or limited to the cortical gray matter and surrounded by a high signal intensity area on FLAIR images. Subcortical infarcts were categorized as such when they do not extend into the cortex and are surrounded by a high signal intensity area on FLAIR images of ≥4 mm in diameter.

Parenchymal defects in the subcortical area without a rim or area of high signal intensity on FLAIR images and without evidence of hemosiderin on the T2*-weighted scan were scored as enlarged perivascular spaces. Enlarged perivascular spaces were excluded from the definition of subcortical infarcts. Enlarged perivascular spaces were documented separately in the whole brain and in the basal ganglia. Cerebellar infarcts were scored without any size criteria. Infarcts covering 2 of the mentioned areas were attributed to the location in which the largest measured diameter was located (in millimeters).⁹

Statistical Analysis

Identification of Subgroups With Different Brain MRI Phenotypes

All brain MRI markers were normalized as z-scores (after multiplication by 100 and natural log transformation when not normally distributed). Variables that were not normally distributed were WMH volumes, solidity, number of WMH, and lateral ventricle volume fraction. Binary variables (presence of microbleeds, infarcts, and enlarged perivascular spaces) were used as −2 and 2 to approximate the z-score distributions of continuous variables. Hierarchical clustering was performed by applying Ward's method in R version 4.1.0 (R Core Team, 2021) and packages factoextra,¹⁶ cluster,¹⁷ and dendextend¹⁸ on 32 brain MRI markers. Hierarchical clustering groups participants together based on similarities in brain MRI markers. The approach starts with every participant as a separate cluster and then repeatedly merging of the 2 closest clusters, subsequently updating the distance matrix. Thus, each cluster is the result of the merge of 2 subclusters, resulting in a hierarchical tree (dendrogram, Figure 2). At each level of the dendrogram, clusters are joined and the number of clusters therefore decreases. This is repeated until only 1 cluster, representing the total group of participants, remains. An optimal number of clusters need to be determined for further analysis. In an optimally clustered data set, the clusters have a high within-cluster cohesion, while having a high separation between different clusters. The optimal dendrogram cut-off, that is, the optimal number of clusters, was determined using the Dunn index (eFigure 2) and the heatmap (Figure 2). The Dunn index is the ratio of the smallest distance between observations in different clusters over the largest between cluster distance and should be maximal. After this procedure, a number of subgroups remained, representing the subgroups with different brain MRI phenotypes.

Figure 2. Heatmap and Dendrogram: Results of the Hierarchical Clustering Analysis.

Red indicates high values, and blue indicates low values. Every line represents a participant, and every column represents a MRI marker. 1: total WMH volume; 2: PV/C WMH volume; 3: PV/C WMH volume frontal lobe; 4: other PV/C WMH volume; 5: PV/C WMH fractal dimension; 6: PV/C WMH volume parietal lobe; 7: PV/C WMH volume temporal lobe; 8: PV/C WMH concavity index; 9: % lateral ventricle volume; 10: PV/C WMH volume occipital lobe; 11: deep WMH volume; 12: deep WMH volume frontal lobe; 13: number of deep WMH; 14: deep WMH volume parietal lobe; 15: other deep WMH volume; 16: deep WMH volume temporal lobe; 17: % total brain volume; 18: % gray matter volume; 19: PV/C convexity; 20: % white matter volume; 21: deep WMH eccentricity; 22: deep WMH fractal dimension; 23: PV/C solidity; 24: deep WMH occipital lobe; 25: number of PV/C; 26: enlarged perivascular spaces in and around basal ganglia; 27: microbleeds; 28: cortical infarcts; 29: subcortical infarcts; 30: enlarged perivascular spaces in the white matter; 31: infarcts (whole brain); 32: cerebellar infarcts. Other PV/C WMH volume or other deep WMH volume were defined as Lesions or parts of lesions that were outside of the brain lobe masks, for example the brain stem and within the internal capsule. PV/C = periventricular/confluent; S = subgroup; WMH = white matter hyperintensity.

Brain MRI markers and cardiovascular risk factors were compared between subgroups with χ² test for binary variables and 1-way analysis of variances for continuous variables by using SPSS version 25 (Chicago, IL). For these analyses, WMH volumes, number of WMH, solidity, ventricle volume fraction, and time to follow-up were log transformed due to a non-normal distribution. A p value <0.05 was considered statistically significant.

Sensitivity Analysis

To assess the robustness of the hierarchical clustering model, we reran the analysis with 2 random subsets of this data set.

Long-Term Outcome Assessment

A Cox proportional hazard model was used to estimate the risk of future dementia occurrence within the brain MRI phenotype subgroups (adjusted for age, sex, and cognitive status at baseline [mild cognitive impairment or normal cognition]). The reference subgroup was chosen based on having the fewest brain abnormalities. SPSS version 25 was used for statistical testing.

Data Availability

The AGES I-II data set cannot be made publicly available because the informed consent signed by the participants prohibits data sharing on an individual level, as outlined by the study approval by the Icelandic National Bioethics Committee. Requests for these data may be sent to the AGES Reykjavik Study Executive Committee, contact: Ms. Camilla Kritjansdottir, Camilla@hjarta.is.

Results

The total sample included 3,056 community-dwelling older adults. The average time to follow-up for dementia outcome (yes/no) was 9.9 ± 2.6 years (min–max range 0.6–13.4 years). The baseline characteristics of the total sample are presented in Table 1. A hierarchical clustering model was applied on brain MRI markers (WMH volumes, brain volumes, WMH shape markers, infarcts, enlarged perivascular spaces, microbleeds). The optimal cut-off of the hierarchical clustering model was determined to be at 15 subgroups (Figure 2), based on the Dunn index (eFigure 2) and the heatmap (Figure 2). The sizes of the subgroups ranged from 42 to 425 participants. As a sensitivity analysis to test the robustness of the clustering method, we reran the model on 2 random subsets (subset 1: n = 2,311 and subset 2: n = 2,250). On average, 68% of participants remained in the same cluster compared with the main analysis. Baseline characteristics of the study sample per subgroup are shown in Table 2. The subgroups differed significantly in age (min–max range of mean age 71.6–78.8 years), sex (min–max range of mean sex distribution 37%–81% females), cognitive status at baseline (min–max range of mean prevalence 3%–15%), hypertension (min–max range of mean prevalence 64%–90%), type 2 diabetes mellitus (min–max range of mean prevalence 4%–15%), cholesterol levels (min–max range of mean 5.20–5.97 mmol/L), coronary artery disease (min–max range of mean prevalence 7%–30%), and time to follow-up (min–max range of mean time to follow-up 8.2–10.9 years; Table 2).

Table 1.

Baseline Characteristics of the Total Study Sample

	Community-dwelling individuals (n = 3,056)
Age (y)	75.6 ± 5.2
Time to follow-up (y)	9.9 ± 2.6
Female sex	1,884 (62)
Mild cognitive impairment	257 (8)
BMI (kg/m2)	26.97 ± 4.19
Hypertension	2,380 (78)
Type 2 diabetes	277 (9)
Cholesterol (mmol/L)	5.71 ± 1.14
Smoking status
Never	1,397 (46)
Former	1,334 (44)
Current	324 (11)
Coronary artery disease	497 (16)

Abbreviation: BMI = body mass index.

Data are shown as means ± SD or percentages of individuals per subgroup.

Table 2.

Baseline Characteristics of the Study Sample per Subgroup

	S1 N = 368	S2 N = 161	S3 N = 425	S4 N = 239	S5 N = 171	S6 N = 333	S7 N = 98	S8 N = 42	S9 N = 105	S10 N = 241	S11 N = 240	S12 N = 87	S13 N = 245	S14 N = 190	S15 N = 111	p Value
Age (y)	74.4 ± 4.7	77.5 ± 5.3	77.8 ± 4.9	73.4 ± 4.1	75.6 ± 5.1	74.3 ± 4.8	74.2 ± 4.8	72.7 ± 3.2	76.8 ± 4.8	71.6 ± 3.4	76.5 ± 5.0	78.2 ± 4.9	75.6 ± 5.0	78.8 ± 5.3	77.3 ± 5.1	<0.001
Time to dementia follow-up (y)	10.33 ± 2.02	8.89 ± 2.80	9.27 ± 2.93	10.89 ± 1.83	9.94 ± 2.33	10.36 ± 2.25	10.41 ± 2.33	10.53 ± 1.66	9.43 ± 2.68	10.70 ± 1.54	9.54 ± 2.89	8.18 ± 3.10	10.28 ± 2.50	8.78 ± 2.85	9.14 ± 2.77	<0.001
Female sex	73	52	50	74	65	54	70	62	54	81	61	37	79	45	43	<0.001
Mild cognitive impairment	7	11	12	3	9	5	3	5	12	5	9	11	8	15	15	<0.001
BMI (kg/m2)	26.74 ± 4.29	27.35 ± 4.29	27.02 ± 4.47	27.18 ± 4.07	26.85 ± 4.11	26.53 ± 4.13	26.99 ± 4.07	27.18 ± 4.28	25.91 ± 3.52	27.18 ± 3.99	26.90 ± 4.19	28.41 ± 4.52	27.29 ± 4.55	27.12 ± 3.51	26.72 ± 3.93	0.021
Hypertension	75	88	81	76	80	70	68	79	84	64	83	90	82	88	82	<0.001
Type 2 diabetes	6	14	10	6	10	6	9	17	10	4	9	11	11	15	15	<0.001
Cholesterol (mmol/L)	5.82 ± 1.10	5.44 ± 1.08	5.67 ± 1.15	5.91 ± 1.16	5.55 ± 1.20	5.77 ± 1.04	5.75 ± 1.10	5.63 ± 0.87	5.69 ± 1.22	5.97 ± 1.07	5.76 ± 1.23	5.52 ± 1.08	5.75 ± 1.14	5.48 ± 1.22	5.20 ± 1.12	0.023
Smoking status
Never	49	35	45	47	44	44	54	36	45	54	49	38	43	44	44	0.034
Former	41	53	47	42	44	44	38	45	44	39	40	47	42	46	45	0.295
Current	9	11	8	10	12	12	8	19	11	7	11	15	15	10	11	0.214
Coronary artery disease	10	24	16	14	19	10	10	7	21	7	21	23	18	29	30	<0.001

Abbreviations: BMI = body mass index; S = subgroup.

Data are shown as means ± SD or percentages of individuals per subgroup. Baseline characteristics and cardiovascular risk factors were compared between subgroups with χ² test for binary variables and 1-way analysis of variances for continuous variables.

Brain MRI markers per subgroup are presented in Table 3. All brain MRI markers differed significantly between subgroups, as could be expected as the subgroups were based on the hierarchical clustering result. The main MRI markers per subgroup are illustrated in a simplified and summarized manner in Figure 3. The brain MRI phenotypes of the subgroups ranged from limited burden (subgroup 10), mostly irregular WMH shape and cerebral atrophy (subgroup 12), mostly irregularly shaped WMH and microbleeds (subgroup 9), mostly cortical infarcts and atrophy (subgroup 15), mostly irregularly shaped WMH and cerebral atrophy (subgroup 3) to multiburden subgroups (subgroup 2, subgroup 14). A complete and detailed description of the main MRI markers of each subgroup can be found in the eResults.

Table 3.

Between-Group Differences of Brain MRI Markers

	S1 N = 368	S2 N = 161	S3 N = 425	S4 N = 239	S5 N = 171	S6 N = 333	S7 N = 98	S8 N = 42	S9 N = 105	S10 N = 241	S11 N = 240	S12 N = 87	S13 N = 245	S14 N = 190	S15 N = 111	p Value
Total WMH volume	11.87 ± 4.69	35.85 ± 22.86	29.93 ± 16.76	7.50 ± 3.33	11.76 ± 7.66	8.02 ± 3.72	5.77 ± 2.25	6.07 ± 2.66	22.90 ± 12.89	3.20 ± 1.44	25.93 ± 15.89	50.32 ± 20.71	24.56 ± 9.63	41.11 ± 21.16	22.85 ± 16.85	<0.001
PV/C WMH volume	10.68 ± 4.59	33.76 ± 22.83	28.51 ± 16.81	6.05 ± 3.10	10.28 ± 7.04	7.57 ± 3.60	4.83 ± 2.14	5.18 ± 2.65	21.05 ± 12.86	2.37 ± 1.29	23.93 ± 15.72	46.99 ± 20.77	21.43 ± 9.71	39.19 ± 21.46	21.29 ± 16.60	<0.001
Deep WMH volume	1.18 ± 0.94	2.09 ± 1.52	1.42 ± 0.91	1.45 ± 0.85	1.49 ± 1.69	0.45 ± 0.39	0.94 ± 0.81	0.89 ± 0.75	1.85 ± 1.41	0.83 ± 0.51	2.00 ± 1.27	3.33 ± 1.51	3.13 ± 1.81	1.91 ± 1.36	1.57 ± 1.35	<0.001
Number of PV/C	14.69 ± 5.70	14.86 ± 5.03	14.96 ± 5.51	14.90 ± 5.89	14.71 ± 5.07	15.27 ± 6.19	15.53 ± 6.06	15.36 ± 5.94	14.78 ± 5.57	16.00 ± 5.98	14.93 ± 6.08	15.62 ± 6.12	15.23 ± 5.94	14.78 ± 5.46	14.57 ± 5.41	<0.001
Number of deep WMH	25.40 ± 15.29	44.69 ± 29.64	41.26 ± 25.69	21.23 ± 13.84	28.89 ± 26.73	16.39 ± 13.72	15.06 ± 10.83	14.67 ± 10.29	41.90 ± 22.40	8.20 ± 5.11	46.33 ± 26.79	86.90 ± 40.35	45.49 ± 16.62	46.88 ± 28.29	34.59 ± 23.67	<0.001
PV/C solidity	0.18 ± 0.07	0.15 ± 0.05	0.13 ± 0.05	0.23 ± 0.11	0.21 ± 0.13	0.20 ± 0.11	0.25 ± 0.11	0.25 ± 0.12	0.13 ± 0.05	0.39 ± 0.16	0.14 ± 0.05	0.03 ± 0.15	0.14 ± 0.04	0.14 ± 0.05	0.17 ± 0.07	<0.001
PV/C convexity	1.11 ± 0.12	0.87 ± 0.18	0.89 ± 0.14	1.13 ± 0.13	1.08 ± 0.14	1.13 ± 0.12	1.13 ± 0.11	1.15 ± 0.11	0.97 ± 0.16	1.12 ± 0.11	0.92 ± 0.19	0.67 ± 0.13	0.97 ± 0.14	0.81 ± 0.16	1.00 ± 0.17	<0.001
PV/C concavity index	1.22 ± 0.07	1.42 ± 0.14	1.41 ± 0.11	1.17 ± 0.07	1.22 ± 0.11	1.19 ± 0.08	1.15 ± 0.07	1.15 ± 0.06	1.35 ± 0.12	1.09 ± 0.06	1.38 ± 0.15	1.59 ± 0.11	1.34 ± 0.10	1.47 ± 0.13	1.31 ± 0.14	<0.001
PV/C fractal dimension	1.69 ± 0.09	1.84 ± 0.11	1.82 ± 0.08	1.65 ± 0.11	1.67 ± 0.14	1.62 ± 0.11	1.58 ± 0.12	1.58 ± 0.11	1.78 ± 0.10	1.48 ± 0.14	1.80 ± 0.11	1.92 ± 0.07	1.81 ± 0.09	1.87 ± 0.09	1.74 ± 0.12	<0.001
Deep WMH eccentricity	0.56 ± 0.06	0.61 ± 0.06	0.63 ± 0.06	0.61 ± 0.07	0.62 ± 0.08	0.65 ± 0.08	0.62 ± 0.08	0.60 ± 0.10	0.61 ± 0.07	0.64 ± 0.08	0.61 ± 0.06	0.66 ± 0.05	0.58 ± 0.05	0.63 ± 0.06	0.62 ± 0.06	<0.001
Deep WMH fractal dimension	1.79 ± 1.79	1.69 ± 0.12	1.65 ± 0.12	1.71 ± 0.13	1.69 ± 0.13	1.59 ± 0.12	1.69 ± 0.15	1.68 ± 0.12	1.68 ± 0.13	1.72 ± 0.13	1.69 ± 0.13	1.60 ± 0.11	1.77 ± 0.08	1.66 ± 0.12	1.70 ± 0.12	<0.001
% lateral ventricle volume	2.61 ± 0.97	3.23 ± 1.09	3.60 ± 1.15	2.18 ± 0.72	2.63 ± 1.08	2.82 ± 0.93	2.19 ± 0.72	2.21 ± 0.91	3.09 ± 1.21	1.85 ± 0.64	2.92 ± 1.06	3.43 ± 0.98	2.50 ± 0.82	3.64 ± 1.21	3.08 ± 1.07	<0.001
% total brain volume	74.52 ± 3.02	71.48 ± 3.59	70.80 ± 3.02	74.27 ± 2.45	72.91 ± 3.07	72.25 ± 2.82	73.44 ± 3.62	74.93 ± 3.16	72.02 ± 3.56	75.73 ± 3.07	71.94 ± 3.60	68.76 ± 4.15	74.55 ± 2.93	70.69 ± 4.01	71.75 ± 3.09	<0.001
% gray matter volume	46.84 ± 2.60	44.50 ± 2.96	43.97 ± 2.62	47.28 ± 2.04	45.95 ± 2.70	45.41 ± 2.43	46.49 ± 3.09	47.86 ± 2.32	45.06 ± 2.85	48.20 ± 2.58	44.96 ± 2.93	42.71 ± 3.40	46.76 ± 2.41	43.72 ± 3.38	44.45 ± 2.83	<0.001
% white matter volume	27.68 ± 1.49	26.98 ± 1.93	26.83 ± 1.35	26.98 ± 1.55	26.96 ± 1.47	26.84 ± 1.49	26.95 ± 1.54	27.07 ± 1.35	26.96 ± 1.74	27.53 ± 1.48	26.98 ± 1.70	26.05 ± 2.03	27.79 ± 1.63	26.97 ± 1.85	27.30 ± 1.59	<0.001
PVS (basal ganglia)	0	9	8	0	5	0	4	0	0	0	8	2	0	6	8	<0.001
PVS	0	0	0	3	10	0	100	0	17	0	99	2	0	17	16	<0.001
Microbleeds	0	14	1	16	12	0	0	5	66	0	2	1	0	9	7	<0.001
Subcortical infarcts	0	100	0	0	8	0	0	100	0	0	0	3	0	22	5	<0.001
Cerebellar infarcts	0	5	0	0	98	0	0	0	0	0	0	0	0	99	29	<0.001
Cortical infarcts	0	4	0	0	1	0	0	0	0	0	0	0	0	4	100	<0.001
Infarcts	0	100	0	0	100	0	0	100	0	0	0	3	0	100	100	<0.001

Abbreviations: PV/C = periventricular/confluent; PVS = enlarged perivascular spaces; S = subgroup; WMH = white matter hyperintensity.

Data shown as mean ± SD or percentages of individuals per subgroup. Brain MRI markers were compared between subgroups with χ² test for binary variables and 1-way analysis of variances for continuous variables. The data for WMH volumes per lobe are presented in eTable 2.

Figure 3. Symbolic Illustration of the Brain MRI Markers per Subgroup.

Small circles indicate low WMH volumes or low amount of cerebral atrophy. Big circles indicate high WMH volumes or cerebral atrophy. WMH shape is illustrated with symbols with 3 varying degrees of shape irregularity (regular, moderate, and irregular). Pie charts indicate percentages of participants with infarcts, enlarged PVS, or microbleeds. Enlarged PVS = enlarged perivascular spaces in and around the basal ganglia; S = subgroup; WMH = white matter hyperintensity.

Subgroup 10 was determined to have the least amount of brain abnormalities and was used as the reference subgroup in the survival analysis (Figure 4). Dementia cases at follow-up ranged from 6% to 46% per subgroup. Compared with the reference subgroup, most other subgroups (except subgroup 4, subgroup 7, and subgroup 8) showed a higher long-term risk for dementia. The range of statistically significant hazard ratios (HRs) across the subgroups varied between 1.01 and 6.18. The subgroup with relatively severe WMH and atrophy (subgroup 12) showed the largest long-term risk for dementia (HR 6.18, 95% CI 3.37–11.32) compared with the reference subgroup (subgroup 10). The multiburden subgroups (subgroup 2: HR 3.68, 2.06–6.61; subgroup 14: HR 3.48, 1.96–6.16), the subgroup with mostly irregularly shaped WMH and cerebral atrophy (subgroup 3: HR 3.33, 1.94–5.73), the subgroup with mostly irregularly shaped WMH and microbleeds (subgroup 9: HR 2.99, 1.60–5.56), and the subgroup with mostly cortical infarcts and atrophy (subgroup 15: HR 3.16, 1.70–5.86) also showed a relatively high long-term risk for dementia.

Figure 4. HRs per Subgroup Based on the Cox Regression Analysis.

Controlled for age, sex, and cognitive status at baseline. Subgroup 10 was used as the reference subgroup in the model, as it was the subgroup with the lowest amount of brain abnormalities. HR = hazard ratio; S = subgroup.

Discussion

We showed that distinct brain MRI phenotypes can be identified in community-dwelling older adults. Some of the 15 distinct subgroups that were identified showed a different long-term dementia risk, with an increased risk, especially in individuals in the multiburden brain pathology subgroups, and in the subgroup with relatively severe WMH and atrophy. Most subgroups showed a significantly increased risk for dementia compared with the reference subgroup that showed the least abnormalities on brain MRI.

The exact underlying structural correlates of the early predisposition to dementia remain largely unknown. Many validated and commonly used brain MRI markers are nonspecific to pathology and disease and commonly occur with aging.¹⁹ Nevertheless, previous studies frequently focused on single or small groups of MRI markers and their relationship with neurovascular or neurodegenerative diseases.^20-23 A combined analysis of brain MRI markers could improve our understanding of the pathophysiology of and early predisposition for dementia, as different brain diseases usually lead to patterns of different brain abnormalities.²⁴ In other fields of research, combined analysis to identify phenotypes has previously been performed. Examples are the identification of asthma phenotypes^25,26 and subphenotypes of chronic obstructive pulmonary disease.^27,28 Moreover, it is frequently used in the field of genetics to identify genotypes, for example, to identify differences in DNA methylation and gene expression in breast cancer.²⁹

There are some previous studies on combined analysis of brain MRI markers or data-driven approaches in the field of dementia research. In a previous study, using an unsupervised deep learning approach on a large diverse data set of T1-weighted brain MRI scans showed that the difference between predicted brain age and chronological age is associated with the presence of different diseases (e.g., schizophrenia, and Alzheimer disease [AD]).³⁰ In another study, a semisupervised deep-clustering method identified 4 neurodegenerative brain MRI patterns based on atrophy regions of interests on T1 scans in a data set including cognitively healthy individuals and patients with cognitive impairment and dementia.³¹ Another study identified 3 brain MRI patterns (neurodegeneration, white matter disease, and typical brain aging) on T1-weighted scans using a machine learning–based method that can be used to identify individual brain health.³² These previous studies used (semi) supervised deep learning approaches, which is a different approach compared with our unsupervised machine learning approach. There are also some previous studies that have used a more similar approach compared with ours, albeit in different patient populations. For example, a previous study has applied hierarchical clustering to identify patterns of markers (including brain MRI markers, blood values, and CSF markers) related to the conversion from mild cognitive impairment to AD.³³ Here, 4 subgroups were identified with a different risk for conversion to AD.³³ The subgroup with the highest risk showed the most severe biomarker profile, for example, the highest WMH volumes, the lowest CSF amyloid beta, the highest CSF tau, and the lowest entorhinal cortical thickness.³³ Another previous study showed that midlife white matter textural properties were associated with future dementia risk.³⁴ A more heterogeneous normal appearing white matter intensity profile was associated with a higher WMH burden in the future, and a more heterogeneous intensity of normal appearing white matter was related to increased dementia risk. Another study found 2 distinct subgroups of mild cognitive impairment based on radiomics similarity networks.³⁵ Significant differences between the 2 mild cognitive impairment subgroups were found, among others, in the regional radiomics similarity networks of the hippocampus, temporal lobe, parahippocampal gyrus, and amygdala, as well as in the gray matter volume and cortical thickness. Furthermore, the 2 subgroups were significantly different from each other in clinical measures and the number of participants progressing to dementia within 3 years.³⁵ Our study is the first to apply an unsupervised machine learning approach in a large group of community-dwelling individuals to assess the association of brain MRI phenotypes and long-term dementia risk. The large sample size of community-dwelling individuals in our study aids the generalizability of our results. Our study is further strengthened by the long follow-up time to the assessment of occurrence of dementia. Because of our study design, in the future, our method could help in assessing an individuals increased dementia risk at an early stage to determine patient prognosis in clinical practice and may aid in patient selection for future treatment studies.

A hierarchical clustering model allows combining of a spectrum of brain MRI markers and to find patterns in these data. We have previously applied the hierarchical clustering method in different data sets with different MRI markers to identify MRI phenotypes of the brain related to future stroke and mortality in patients with manifest arterial disease,⁷ as well as related to increased postoperative delirium risk in preoperative patients.³⁶ This study assessed brain MRI phenotypes in relation to long-term dementia risk in community-dwelling older adults. Our results identified 15 distinct subgroups of individuals with different distributions of brain MRI markers of neurodegenerative and neurovascular disease. The multiburden group with the highest long-term risk for dementia (subgroup 12) does show markers of SVD, such as high WMH volumes and an irregular WMH shape, but includes only few individuals with brain infarcts. In addition, subgroup 12 showed the most severe cerebral atrophy, which may suggest that this subgroup has more underlying neurodegenerative pathology. Subgroup 2 has, similar to subgroup 12, high WMH volumes and an irregular WMH shape but also includes a high number of participants with subcortical, cerebellar, and cortical infarcts. Atrophy is less prominent in subgroup 2 compared with subgroup 12. Subgroup 15 may include mostly patients with large vessel disease, as WMH volumes and shape are only moderately abnormal, while all participants in this group have cortical infarcts. We showed that different brain MRI phenotypes, characterized by a distinct combination of brain MRI markers, predispose to occurrence of dementia and are related to different long-term dementia risks.

Strengths of our study include the use of multiple brain MRI markers in 1 framework, a large sample size, and a long follow-up period for dementia outcome. Furthermore, we mostly included markers that can be (semi) automatically detected on brain MRI scans (e.g., WMH volumes and brain atrophy) and the inclusion of novel brain MRI markers (such as WMH shape). An automated, unsupervised approach to identify groups was applied that allowed us to identify novel patterns of brain MRI markers. Limitations of this study might be the somewhat subjective cut-offs within the model, such as the dendrogram cut-off. However, to increase objectivity, we used the Dunn index and the heatmap to determine the cut-off for the number of subgroups. Another limitation could be that the model is dependent on the selection of brain MRI markers that were included in the model. We therefore chose to include etiologically and prognostically relevant and validated brain MRI markers of which many can be quantified automatically. Moreover, the hierarchical clustering results could also be influenced by the choice of linkage method (e.g., Ward's, centroid), which is used to delineate the subgroups. We have chosen to use Ward's criteria as a linkage method because it generates subgroups with minimal within-subgroup variance and to maximize the between-subgroup variance, which we deemed as most suitable for this data set and type of analysis. Another general limitation of the clustering method could be that in our sensitivity analyses, we showed that there is some dependency of the clustering results based on the number and selection of participants. For future research the Subtype and Stage Inference (SuStaIn) method, an unsupervised machine learning technique that identifies population subgroups with common patterns of disease progression could be an interesting approach.³⁷ SuStaIn could provide additional insights since it combines traditional clustering with disease progression modeling, but the effect of this approach on reproducibility is also of interest. Another limitation of this study could be that most neuroimaging research 1.5T MRI scanners are nowadays replaced with a 3T MRI system, which was not yet the case at the time of the data collection for our study. Nevertheless, we did successfully identify distinct brain MRI phenotypes based on our data set.

In conclusion, distinct brain MRI phenotypes are related to varying long-term risks of developing dementia. Brain MRI phenotypes may assist in an improved understanding of the structural correlates of dementia predisposition. These findings may aid in the future to determine patient prognosis and for patient selection for future treatment studies.

Glossary

AD

Alzheimer disease

AGES

Age-Gene/Environment Susceptibility

FLAIR

fluid-attenuated inversion recovery

HR

hazard ratio

SVD

small vessel disease

WMH

white matter hyperintensity

Appendix. Authors

Name	Location	Contribution
Jasmin Annica Keller, MSc	Department of Radiology, Leiden University Medical Center	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data
Sigurdur Sigurdsson, MSc, PhD	Icelandic Heart Association	Major role in the acquisition of data; drafting/revision of the manuscript for content
Bárbara Schmitz Abecassis, MSc	Department of Radiology, Leiden University Medical Center	Analysis or interpretation of data; drafting/revision of the manuscript for content
Ilse M.J. Kant, MSc, PhD	Clinical Artificial Intelligence Implementation and Research Lab (CAIRELab) and Department of Information Technology & Digital Innovation; Department of Digital Health, University Medical Center Utrecht	Analysis or interpretation of data; drafting/revision of the manuscript for content
Mark A. Van Buchem, MD, PhD	Department of Radiology, Leiden University Medical Center	Drafting/revision of the manuscript for content, including medical writing for content
Lenore J. Launer, PhD	Laboratory of Epidemiology and Population Science, National Institute on Aging	Drafting/revision of the manuscript for content, including medical writing for content
Matthias J.P. van Osch, PhD	Department of Radiology, Leiden University Medical Center	Drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data
Vilmundur Gudnason, MD, PhD	Icelandic Heart Association; Faculty of Medicine, University of Iceland	Drafting/revision of the manuscript for content, including medical writing for content
Jeroen de Bresser, MD, PhD	Department of Radiology, Leiden University Medical Center	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data

Study Funding

The Age, Gene/Environment Susceptibility Reykjavik Study was supported by NIH contracts N01-AG-1-2100 and HHSN27120120022C, the NIA Intramural Research Program, Hjartavernd (the Icelandic Heart Association), and the Althingi (the Icelandic Parliament). This work was supported by an Alzheimer Nederland grant (WE.03-2019-08) to J. de Bresser.

Disclosure

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