Cohort characteristics
The basic clinical and demographic information of the cohort is reported in Table 1. As expected, The AD group significantly differed from the CN group in terms of percentage of APOE ε4 carriers, cognitive measures, and CSF markers. There were no significant differences between our AD and CN groups in terms of sex and age.
FDG-PET Clustering
The overall AD group’s hypometabolism pattern showed neurodegeneration in posterior cingulate and frontal cortical regions and deeper structures such as the hippocampus (Fig. 1a). The clustering model resulted in two main distinguishing patterns, neurodegeneration in cortical versus limbic pathways (Fig. 1c). The Calinski-Harabasz index peaked at five clusters, for the Davies-Bouldin index was lower at five, the Dunn index plateaus at five and the Silhouette index was highest at 3–5 clusters (Suppl. Figure 6). Therefore, we chose five clusters as the optimal solution for the FDG-PET model. The FDG-PET model was split into five distinct hypometabolism-based subtypes. Three subtypes showed cortical-predominant hypometabolism of differing severity and spatial distribution. The Cortical Predominant posterior subtype had cortical hypometabolism mainly in the posterior regions (9%), whereas the Cortical Predominant and Cortical Predominant + subtypes showed more widespread cortical hypometabolism (32% and 11% respectively). The Cortical Predominant + subtype had greater hypometabolism than the other two cortical predominant subtypes. Although all the subtypes showed some hypometabolism in the hippocampus, these subtypes had proportionally less involvement of this region compared to the cortical areas. Two subtypes displayed limbic hypometabolism, focal to the medial temporal and deeper structures (amygdala, hippocampus). Here, a principal Limbic Predominant subtype (36%) could be distinguished from and a Limbic Predominant frontal subtype (13%). In the clustering dendrogram (Fig. 1), the Limbic Predominant frontal cluster originates from its own branch, whereas the Limbic Predominant cluster comes from the same branch as the Cortical Predominant posterior cluster. Thus, the clustering separates these subtypes by a frontal versus a posterior hypometabolism pattern. By contrast, the Cortical Predominant and Cortical Predominant + clusters are separated on the opposite side of the dendrogram by the severity of their cortical hypometabolic patterns.
For inter-modality comparisons, the corresponding atrophy patterns in these subtypes were plotted (Fig. 1e). Based on visual comparison of the w-scores, the brain maps were topographically similar across FDG-PET subtypes, showing AD-typical atrophy in medial temporal, hippocampal, and some frontal areas. However, the atrophy pattern of the Cortical Predominant subtype was not as widespread in the cortical regions compared to the hypometabolism. Additional maps using PVC SUVRs from PETSurfer were plotted (Suppl. Figure 8). These brain maps did not differ greatly from our maps in Fig. 1 topographically but did result in lower w-scores.
Regarding demographic and clinical differences among the FDG-PET AD subtypes (Table 2), the Cortical Predominant + subtype was the youngest (67.5 years), had the earliest age at onset (64.7 years), more pronounced language impairment, and lowest executive function scores. This subtype also had the highest grey matter volume-based and SUVR-based hippocampus-to-cortex ratios. The other two cortical subtypes (Cortical Predominant, Cortical Predominant posterior) had a higher SUVR-based hippocampus-to-cortex ratio than the limbic subtypes. Cortical Predominant posterior also had a high hippocampus-to-cortex ratio using grey matter volumes compared to the limbic subtypes. Among the limbic subtypes, Limbic Predominant frontal was the oldest, had latest age at onset and worst language scores. There were no significant differences between the subtypes for the other variables: sex, disease duration, years of education, APOE ε4 carriers, MMSE, CDR, cognitive measures of memory, and CSF biomarkers. Although not statistically significant, two of the cortical subtypes, Cortical Predominant and Cortical Predominant+, had lower percentage of APOE ε4 carriers and Limbic Predominant frontal the highest percentage of APOE ε4carriers.
Table 2
Demographic and clinical characteristics of the FDG-PET AD subtypes
Demographic & Clinical Characteristics
|
Cortical Predominant
|
Cortical Predominant+
|
Limbic Predominant
|
Limbic Predominant frontal
|
Cortical Predominant posterior
|
Cognitively Normal
|
p values
|
N (%)
|
57 (32%)
|
20 (11%)
|
64 (36%)
|
23 (13%)
|
16 (9%)
|
176
|
―
|
Women (%)
|
22 (39%)
|
10 (50%)
|
32 (50%)
|
10 (43%)
|
7 (44%)
|
84 (48%)
|
0.768
|
Age (years)
|
75 (7.9)
|
67 (8.8)
|
74 (6.8)
|
78 (6.6)
|
73 (8.8)
|
75 (7.1)
|
< 0.001
|
Disease Duration (years)
|
2.5 (2.7)
|
2.7 (2.1)
|
2.9 (2.5)
|
3.1 (2.9)
|
1.8 (2.1)
|
-
|
0.274
|
Age at Onset (years)
|
72 (8.1)
|
65 (8.3)
|
71 (7)
|
75 (7.4)
|
71 (8.8)
|
-
|
< 0.001ab
|
Education (years)
|
15 (2.5)
|
16 (2.8)
|
16 (2.9)
|
15 (3.5)
|
15 (2.2)
|
17 (2.6)
|
0.361
|
APOE ɛ4 (%)
|
35 (61%)
|
11 (55%)
|
48 (75%)
|
20 (87%)
|
12 (75%)
|
34 (19%)
|
0.083
|
MMSE
|
23 (2.1)
|
22 (2.6)
|
24 (2)
|
23 (1.9)
|
23 (1.7)
|
29 (1.4)
|
0.167
|
Global CDR
|
0.77 (0.25)
|
0.88 (0.22)
|
0.73 (0.25)
|
0.77 (0.25)
|
0.89 (0.4)
|
0.025 (0.16)
|
0.225
|
ADNI-EF
|
-0.43 (0.62)
|
-0.99 (0.73)
|
-0.19 (0.61)
|
-0.6 (0.51)
|
-0.58 (0.7)
|
0.8 (0.47)
|
< 0.001bc
|
ADNI-MEM
|
-0.85 (0.32)
|
-0.89 (0.32)
|
-0.66 (0.34)
|
-0.93 (0.31)
|
-0.8 (0.4)
|
0.89 (0.52)
|
0.007
|
ADNI-LAN
|
-0.24 (0.53)
|
-0.43 (0.52)
|
0.041 (0.54)
|
-0.5 (0.61)
|
-0.31 (0.56)
|
0.82 (0.52)
|
< 0.001bce
|
CSF t-tau (pg/ml)
|
378 (145)
|
407 (143)
|
415 (167)
|
335 (98)
|
335 (152)
|
223 (70)
|
0.958
|
CSF p-tau (pg/ml)
|
38 (15)
|
41 (16)
|
43 (19)
|
33 (10)
|
33 (15)
|
20 (6.3)
|
0.961
|
CSF Aβ (pg/ml)
|
585 (248)
|
588 (173)
|
588 (169)
|
596 (150)
|
578 (139)
|
1204 (333)
|
0.537
|
SUVR-based Hippocampus-to-cortex ratio
|
0.35 (0.04)
|
0.38 (0.047)
|
0.3 (0.034)
|
0.32 (0.032)
|
0.33 (0.032)
|
0.32 (0.029)
|
< 0.001
|
Grey Matter Volume-based Hippocampus-to-cortex ratio
|
0.16 (0.021)
|
0.19 (0.029)
|
0.16 (0.02)
|
0.16 (0.023)
|
0.17 (0.025)
|
0.18 (0.021)
|
< 0.001fg
|
Total Grey Matter Volume (mm³)
|
558154 (33242)
|
544820 (35401)
|
554040 (38026)
|
537830 (47861)
|
539736 (36515)
|
580718 (40407)
|
0.129
|
Total Average Cortical Uptake (SUVR)
|
1.5 (0.16)
|
1.5 (0.12)
|
1.6 (0.18)
|
1.6 (0.11)
|
1.6 (0.2)
|
1.7 (0.19)
|
0.021
|
The values shown in the table are the means with standard deviations in brackets except for number of individuals, women and APOE ε4 for which the percentages are provided. The reported p-values correspond to Χ2 tests which were used for categorical variables and Kruskal-Wallis for continuous variables and were corrected for multiple comparison with the Holm-Šidák method. Footnotes indicate cases where p values were significant in the post hoc pairwise comparisons across AD subtypes, p < 0.05. The CN group data is displayed for reference. Abbreviations: CP: Cortical Predominant, CP+: Cortical Predominant+, LP: Limbic Predominant, LP fr.: Limbic Predominant frontal, CP post.: Cortical Predominant posterior, CN: cognitively normal individuals, APOE ε4: apolipoprotein E ε4 allele, MMSE: Mini Mental State Examination, CDR: The Clinical Dementia Rating Scale, ADNI-EF Alzheimer’s Disease Neuroimaging Initiative executive function composite score, ADNI-MEM: Alzheimer’s Disease Neuroimaging Initiative memory composite score, ADNI-LAN: Alzheimer’s Disease Neuroimaging Initiative language composite score, CSF t-tau: total tau, CSF p-tau: phosphorylated tau, CSF Aβ: amyloid-beta 1-42 peptide.
- Cortical Predominant+ < Limbic Predominant, p < 0.05
- Cortical Predominant+ < Cortical Predominant, p < 0.05
- Limbic Predominant < Limbic Predominant frontal, p < 0.05
- Limbic Predominant frontal < Limbic Predominant, p < 0.05
- Limbic Predominant < Cortical Predominant+, p < 0.05
- Limbic Predominant frontal < Cortical Predominant+, p < 0.05
- Limbic Predominant < Cortical Predominant, p < 0.05
- Limbic Predominant < Cortical Predominant posterior, p < 0.05
- Cortical Predominant posterior < Cortical Predominant+, p < 0.05
- Cortical Predominant < Cortical Predominant+, p < 0.05
MRI Clustering
The overall AD group showed atrophy in the expected medial temporal regions such as hippocampus and amygdala (Fig. 1b), which will be referred to as the a ‘typical’ AD pattern for MRI. Similar to FDG-PET subtypes, clustering revealed a distinction between either a limbic or a cortical pathway in MRI (Fig. 1d). The Calinski-Harabasz index peaked at three but was still high at five clusters, the Davies-Bouldin index was lower at five, the Dunn index was high for five albeit plateaued at six before a sharp increase after that and the Silhouette index was highest at three to six clusters (Suppl. Figure 7). Therefore, we chose five clusters as the optimal solution for the MRI model based on these results and considering prior work identifying five biological subtypes. Another reason for choosing a higher cluster solution was based on the lack of sensitivity for finding atypical patterns when implementing a three- and four- cluster solutions. The MRI clustering model was split into five atrophy-based subtypes. In contrast to the FDG-PET subtypes which were limited to cortical and limbic subtypes, the MRI subtypes showed additional ‘minimal’ versus ‘diffuse’ atrophy patterns. Similar to the FDG-PET Cortical Predominant subtypes, a Cortical Predominant MRI subtype (19%) showed greater cortical atrophy relative to the hippocampus. The Limbic Predominant subtype (27%) had the opposite pattern with greater atrophy in the hippocampus relative to the cortex. The Minimal subtype (19%) had some atrophy in the hippocampus and amygdala, but very little atrophy compared to cognitively normal individuals in the cortical regions. There were two diffuse atrophy subtypes, one with greater overall atrophy, Diffuse+ (6%), and one with similarly diffuse but less severe atrophy (28%).
Based on the inter-modality comparison, the atrophy-based subtypes displayed hypometabolism of differing severity in temporo-parietal and lateral temporal regions often described to be the ‘typical AD’ pattern in FDG-PET scans (Fig. 1f). Compared to the corresponding atrophy maps of the FDG-PET subtypes (Fig. 1e), the hypometabolism maps (Fig. 1f, Suppl. Figure 9) were more topographically similar to the MRI subtypes when based on visual comparison of w-scores. These corresponding maps showed both more pronounced (higher w-scores) and more widespread hypometabolism in the Minimal and Cortical Predominant subtypes compared to their atrophy maps (Fig. 1f, Suppl. Figure 9).
Regarding demographic and clinical differences, among the MRI AD subtypes (Table 3), the Diffuse subtype had the lowest executive function scores compared to the Minimal and Limbic Predominant subtypes. Diffuse, Diffuse+, and Cortical Predominant had significantly worse executive function scores compared to the Minimal subtype. Minimal and Limbic Predominant subtypes had significantly lower SUVR-based hippocampus-to-cortex ratios to Cortical Predominant. Significant differences were also found in the grey matter volume-based hippocampus-to-cortex ratios: Cortical Predominant had the highest hippocampus-to-cortex ratio. There were no significant differences between the subtypes for the other variables: sex, age, disease duration, age at onset, years of education, APOE ε4 carriers, MMSE, CDR, cognitive measure of memory and language, and CSF biomarkers. Despite not showing a significant difference, Diffuse + had the highest proportion of APOE ε4 carriers (90%) and was the oldest group (79.3 years).
Table 3: Demographic and clinical characteristics of the MRI AD subtypes.
Demographic & Clinical Characteristics
|
Cortical Predominant
|
Diffuse
|
Limbic Predominant
|
Diffuse+
|
Minimal
|
Cognitively Normal
|
p values
|
N (%)
|
35 (19%)
|
51 (28%)
|
49 (27%)
|
10 (6%)
|
35 (19%)
|
176
|
―
|
Women (%)
|
10 (29%)
|
30 (59%)
|
22 (45%)
|
6 (60%)
|
13 (37%)
|
84 (48%)
|
0.049
|
Age (years)
|
70 (9.5)
|
74 (8.4)
|
76 (6.3)
|
79 (4.9)
|
73 (6.7)
|
75 (7.1)
|
0.004
|
Disease Duration (years)
|
2.2 (2.3)
|
3.4 (2.5)
|
2.5 (2.7)
|
4.4 (3.2)
|
1.9 (2.1)
|
―
|
0.002
|
Age at Onset (years)
|
68 (9.4)
|
71 (8.6)
|
74 (6.7)
|
75 (7.3)
|
71 (6.7)
|
―
|
0.021
|
Education (years)
|
15 (3)
|
15 (2.9)
|
16 (2.6)
|
16 (2.7)
|
16 (2.6)
|
17 (2.6)
|
0.197
|
APOE ɛ4 (%)
|
22 (63%)
|
38 (75%)
|
31 (63%)
|
9 (90%)
|
26 (74%)
|
34 (19%)
|
0.329
|
MMSE
|
23 (2.4)
|
23 (2)
|
23 (2.1)
|
23 (2.1)
|
24 (2.1)
|
29 (1.4)
|
0.767
|
Global CDR
|
0.79 (0.25)
|
0.83 (0.24)
|
0.74 (0.32)
|
0.89 (0.22)
|
0.72 (0.25)
|
0.025 (0.16)
|
0.121
|
ADNI-EF
|
-0.6 (0.62)
|
-0.72 (0.6)
|
-0.32 (0.5)
|
-0.63 (0.72)
|
-0.0096 (0.74)
|
0.8 (0.47)
|
<0.001[a][b][c]
|
ADNI-MEM
|
-0.78 (0.32)
|
-0.91 (0.33)
|
-0.78 (0.29)
|
-0.85 (0.43)
|
-0.65 (0.4)
|
0.89 (0.52)
|
0.059
|
ADNI-LAN
|
-0.22 (0.41)
|
-0.4 (0.52)
|
-0.22 (0.55)
|
-0.13 (0.67)
|
0.11 (0.68)
|
0.82 (0.52)
|
0.007
|
CSF t-tau (pg/ml)
|
390 (164)
|
398 (145)
|
380 (141)
|
280 (77)
|
377 (161)
|
223 (70)
|
0.065
|
CSF p-tau (pg/ml)
|
39 (17)
|
40 (16)
|
37 (14)
|
26 (7.4)
|
39 (18)
|
20 (6.3)
|
0.041
|
CSF Aβ (pg/ml)
|
564 (168)
|
575 (159)
|
661 (245)
|
536 (128)
|
551 (181)
|
1204 (333)
|
0.425
|
SUVR-based Hippocampus-to-cortex ratio
|
0.35 (0.047)
|
0.34 (0.048)
|
0.32 (0.038)
|
0.34 (0.053)
|
0.31 (0.043)
|
0.32 (0.029)
|
<0.001[d][e]
|
Grey Matter Volume-based Hippocampus-to-cortex ratio
|
0.18 (0.021)
|
0.17 (0.026)
|
0.15 (0.018)
|
0.16 (0.027)
|
0.15 (0.022)
|
0.18 (0.021)
|
<0.001p[f][g]
|
Total Grey Matter Volume (mm³)
|
561400 (17261)
|
518539 (12671)
|
552539 (15299)
|
478411 (11842)
|
606360 (16474)
|
580718 (40407)
|
<0.001ms[h]
|
Total Average Cortical Uptake (SUVR)
|
1.6 (0.16)
|
1.5 (0.14)
|
1.6 (0.15)
|
1.4 (0.081)
|
1.7 (0.21)
|
1.7 (0.19)
|
<0.001a[i][j][k]
|
The values shown in the table are the means with standard deviations in brackets except for number of individuals, women and APOE ε4 for which the percentages are provided. Χ2 tests were used for categorical variables and Kruskal-Wallis for continuous variables and were corrected for multiple comparison with the Holm-Šidák method. Footnotes indicate cases where p values were significant in the post hoc pairwise comparisons across AD subtypes, p < 0.05. The CN group data is displayed for reference. Abbreviations: CP: Cortical Predominant, LP: Limbic Predominant, Min: Minimal atrophy, Dif: Diffuse, Dif+: Diffuse+, CN: cognitively normal individuals, APOE ε4: apolipoprotein E ε4 allele, MMSE: Mini Mental State Examination, CDR: The Clinical Dementia Rating Scale, ADNI-EF Alzheimer’s Disease Neuroimaging Initiative executive function composite score, ADNI-MEM: Alzheimer’s Disease Neuroimaging Initiative memory composite score, ADNI-LAN: Alzheimer’s Disease Neuroimaging Initiative language composite score, CSF t-tau: total tau, CSF p-tau: phosphorylated tau, CSF Aβ: amyloid-beta 1-42 peptide.
- Diffuse < Minimal, p < 0.05
- Cortical Predominant < Minimal, p < 0.05
- Diffuse < Limbic Predominant, p < 0.05
- Minimal < Cortical Predominant, p < 0.05
- Limbic Predominant < Cortical Predominant, p < 0.05
- Diffuse < Cortical Predominant, p < 0.05
- Minimal < Diffuse, p < 0.05
- Limbic Predominant < Minimal, p < 0.05
- Diffuse+ < Minimal, p < 0.05
- Diffuse+ < Cortical Predominant, p < 0.05
- Diffuse+ < Limbic Predominant, p < 0.05
Individual-level Subtype Allocations
To assess the consistency between the two modalities in subtype assignments, the subtype categorizations for individuals were compared across both FDG-PET and MRI (Fig. 2). We propose that the cortical subtypes and limbic subtypes are most similar between the FDG-PET and MRI subtypes. Namely, the FDG-PET Cortical Predominant, Cortical Predominant posterior and Cortical Predominant + subtypes are equivalent with MRI Cortical Predominant, Diffuse or Diffuse + patterns topographically. Whereas FDG-PET Limbic Predominant and Limbic Predominant frontal are equivalent to MRI Limbic Predominant. As a Minimal pattern was only found in MRI, we do not think that this subtype has an equivalent in FDG-PET.
The agreement between the FDG-PET and MRI subtype allocations was low as this was less than 50%. Although the compared subtypes showed similar topographies of neurodegeneration (i.e., cortical/limbic predominant hypometabolism and atrophy, respectively) they did not match at the individual-level. All possible combinations of allocated FDG-PET and MRI subtypes of varying percentages were found (Fig. 2a, b). AD individuals classified into the FDG-PET Cortical Predominant subtype matched most with the MRI Limbic Predominant (33.3%) and MRI Cortical Predominant (24.6%) subtypes. By contrast, individuals classified as FDG-PET Cortical Predominant + matched best with MRI Cortical Predominant (35%) and Diffuse (35%). FDG-PET Limbic Predominant best matched with three MRI subtypes: Limbic Predominant (28.1%), Minimal (26.6%) and Diffuse (25%). FDG-PET Limbic Predominant frontal best matched with MRI Diffuse (52.2%). FDG-PET Cortical Predominant posterior matched best with MRI Limbic Predominant (37.5%).
Data-driven versus Hypothesis-driven Subtyping
Clustering-based and prior hypothesis-based subtypes were compared within the framework of typicality and severity2 in both FDG-PET and MRI (Fig. 3). Within each modality, data-driven and hypothesis-driven subtypes overlapped with each other reasonably well for most subtypes (Fig. 3a, b). The agreement between MRI data-driven and MRI hypothesis-driven limbic predominant subtypes (55.6%) was better than that between cortical predominant subtypes (30%) (Fig. 3b). Contrarily, the agreement between FDG-PET data-driven and FDG-PET hypothesis-driven cortical predominant subtypes (90%) was better than that between limbic predominant subtypes (82%) (Fig. 3a).
Association between typicality and severity by modality differed. Correlation between typicality and severity (Table 4) was significant in FDG-PET (r2 = 0.25), but not in MRI. The severity measures (r2 = 0.15) in FDG-PET and MRI were more strongly associated with each other than typicality measures (r2 = 0.02). Additionally, FDG-PET subtypes are more separable across the typicality axis than MRI, this is evident when comparing averaged normalised values of typicality and severity for each subtype (Fig. 4). Whereas for MRI subtypes, there was a clearer split of the along the severity axis.
Table 4
Correlations between Typicality & Severity in FDG-PET and MRI.
Model
|
R²
|
P value
|
FDG-PET Typicality and FDG-PET Severity
(Fig. 3a)
|
0.25
|
< 0.001**
|
MRI Typicality and MRI Severity (Fig. 3b)
|
0.0011
|
0.53
|
FDG-PET Severity and MRI Severity
|
0.15
|
< 0.001**
|
FDG-PET Typicality and MRI Typicality
|
0.02
|
0.0073*
|
Pearson correlations between measures of ‘typicality’ and ‘severity’ in both modalities. |
* P < 0.05. |
** P < 0.001. |