Atrophy subtypes and the ATN classication scheme in Alzheimer’s disease

BACKGROUND We investigated the association between atrophy subtypes of Alzheimer’s disease (AD), the ATN classication scheme, and key demographic and clinical factors, in two cohorts with different source characteristics (a highly selective research-oriented cohort, ADNI; and a naturalistic heterogeneous clinically-oriented cohort, Karolinska Imaging Dementia Study (KIDS). METHODS A total of 382 AD patients were included. Factorial analysis of mixed data was used to investigate associations between AD subtype based on brain atrophy patterns, ATN proles based on cerebrospinal uid biomarkers, and age, sex, Mini Mental State Examination (MMSE), cerebrovascular disease (CVD) (burden of white matter signal abnormalities, WMSA), and APOE genotype.


Background
Disentangling the heterogeneity within Alzheimer's disease (AD) has become an important task in order to guide personalized interventions [1,2]. Neuropathological and neuroimaging studies have consistently identi ed three biological subtypes of AD: typical, limbic-predominant, and hippocampal-sparing AD.
Typical AD is characterised by a balanced count of neuro brillary tangles (NFT) or atrophy in hippocampus and association cortex. Limbic-predominant AD has NFT or atrophy predominantly in the hippocampus. Hippocampal-sparing AD has NFT or atrophy predominantly in the association cortex.
Several neuroimaging studies have also identi ed a fourth subtype with minimal signs of brain atrophy, i.e., the minimal atrophy AD subtype [3,4]. However, very few studies have investigated the pathophysiological background of these subtypes in vivo [5], which is needed to completely disentangle the biological heterogeneity within AD.
Another way to stratify AD patients and also inform on their pathophysiological background is the ATN classi cation scheme, which is based on dichotomous categories (normal/abnormal) of amyloid-beta (A), tau (T), and neurodegeneration (N) biomarkers. To our knowledge, only one study investigated AD subtypes in combination with ATN pro les, and that study was performed in mild cognitive impairment (MCI) patients [6].
A task that remains to be done is the incorporation of a category for cerebrovascular (CVD, V) to the ATN scheme [7]. White matter signal abnormalities (WMSA) on magnetic resonance imaging (MRI) are a wellestablished marker of CVD. WMSA are implicated in AD pathogenesis [8,9] and are commonly found in cognitively unimpaired older individuals [10,11]. Including the V category in the scheme is important to advance our understanding of associations between ATV pathologies and their contribution to neurodegeneration. Stratifying AD patients into biological subtypes extends the N category by including a topographical dimension. The topographical dimension likely corresponds to different combinations of ATV and demographic, clinical, and genetic factors. A recent conceptual framework proposed how all these factors interrelate with each other giving rise to the biological subtypes of AD [5]. However, this framework has not been tested empirically.
The aim of this study was to investigate the association between AD subtypes, ATN pro les, and key demographic and clinical factors. We evaluated AD subtypes in combination with ATN pro les in two cohorts: a homogeneous research-oriented cohort (the ADNI study: Alzheimer's Disease Neuroimaging Initiative), and a heterogeneous clinically oriented cohort (the KIDS study: Karolinska Imaging Dementia Study). Investigating AD subtypes and ATN pro les in cohorts with different characteristics is relevant because these subtypes are thought to result from risk factors, protective factors, and comorbid brain pathologies [5] that are differently represented in research-and clinically-oriented cohorts 12 . We hypothesized that the distribution of AD subtypes and ATN pro les would differ depending on cohort and demographic and clinical characteristics. We hypothesized that older patients would include a higher proportion of women with higher WMSA burden, higher proportion of A + T + N + individuals, and lower global cognitive performance; all these related with a higher proportion of individuals classi ed with typical or limbic-predominant AD subtypes. Younger patients would include a higher proportion of men with lower WMSA burden.
WMSA were investigated as a marker of CVD. In the ADNI cohort, WMSA were assessed through automatically segmented white matter hypointensities from FreeSurfer 6.0.0 (https://surfer.nmr.mgh.harvard.edu/). FreeSurfer is increasingly used to automatically segment WMSA in the form of hypointensities in the T1-weighted MRI sequence [20][21][22][23][24]. In KIDS, the MRI data is clinical and there is variation across patients in scanning parameters. In such data, the Fazekas visual rating scale [25] is appropriate as a measure of WMSA because it is not in uenced by variation in scanning parameters [26]. Hence, in the KIDS cohort, WMSA were assessed through the Fazekas scale on FLAIR images. Importantly, hypointense WMSA from FreeSurfer and the Fazekas scale are strongly associated with each other [27]. We followed a previous study that demonstrated that hypointense WMSA from FreeSurfer and Fazekas scores can be combined together by converting hypointense WMSA into low and high CVD burden [27]. Hence, we computed a unique WMSA variable by applying a cut-off of 0.00321 on hypointense WMSA from FreeSurfer after total intracranial volume (TIV) adjustment. This creates two categories of low and high CVD burden that are analogous to Fazekas scores of 0 or 1 de ned as low WMSA and Fazekas scores of 2 or 3 de ned as high WMSA [27]. Regional brain atrophy was assessed with visual rating scales as detailed elsewhere [4,28]. Brie y, medial temporal atrophy (MTA) was assessed with the Scheltens' scale [29], posterior atrophy (PA) with the Koedam's scale [30], and atrophy in the frontal lobe with the global cortical atrophy scale -frontal 2.4. AD subtypes based on patterns of brain atrophy Deviation from normality in visual ratings was determined using previously published cut-offs [28]. The MTA scores ≥ 1.5, ≥ 1.5, ≥ 2, ≥2.5 were considered abnormal for the respective age ranges 45-64, 65-74, 75-84, and 85-94 years. Since an age-correction does not improve PA and GCA-F diagnostic performance [28], a score ≥ 1 was considered abnormal irrespectively of the age range [28]. The three AD subtypes identi ed in the previous literature [32,33] were de ned based on the combination of MTA, PA, and GCA-F, as in previous studies [4,6,34,35]. The minimal atrophy AD subtype [3,34,36]

Statistical analysis
Page 6/23 The main aim of this study was to investigate the association between AD subtypes, ATN pro les, and key demographic and clinical factors. Given the nature of our data, which included both continuous and categorical variables, we applied a multivariate method for data analysis called factorial analysis of mixed data (FAMD) [37]. The main strength of FAMD is that it accommodates both quantitative and qualitative data simultaneously. FAMD works as a principal component analysis for quantitative data and as a multiple correspondence analysis for qualitative data [37]. In our FAMD model, age and MMSE scores were included as continuous variables, and the cohort (ADNI vs. KIDS), ATN categories, AD subtypes, sex (men vs. women), and WMSA burden (high vs. low) were included as categorical variables. A complementary FAMD model was conducted by adding APOE genotype as a categorical variable (carriers of at least one ε4 allele vs. non-carriers). One-way ANOVA was used for continuous variables and the chi-square test was used for categorical data. Missing data on MMSE was estimated via the MissForest algorithm [38] for six KIDS patients. All statistical analyses were conducted using the R statistical software (R Foundation for Statistical Computing, Vienna, http://www-R-project.org). A p-value ≤0.05 was deemed statistically signi cant.

Cohort characteristics
Cohort characteristics are shown in Table 1a (N = 382). ADNI patients were signi cantly older with higher scores in MMSE and a lower frequency of women as compared with KIDS patients. Further, ADNI patients showed a signi cantly higher WMSA burden as well as a higher frequency of abnormal CSF Aβ42 and phosphorylated tau levels, while KIDS patients showed a signi cantly higher frequency of abnormal CSF total tau levels. Due to the reduced number of amyloid-beta negative (A-) patients (N = 79), A-groups were excluded from subsequent analyses. The amyloid-beta positive (A+) subsample is shown in Table 1b (N = 303). In the A + subsample, ADNI patients were signi cantly older with higher scores in MMSE, a higher WMSA burden, and a lower frequency of women and abnormal CSF phosphorylated tau levels, as compared with KIDS patients. Visual inspection of the data shows that typical AD was the most frequent subtype in both ADNI and KIDS (Fig. 2). Limbic-predominant and minimal atrophy AD were more frequent in KIDS, and hippocampal-sparing AD was more frequent in ADNI (Fig. 2). Minimal atrophy AD patients were signi cantly younger than patients from the other subtypes and had a lower WMSA burden than typical AD patients (Table 2). Typical AD patients had worse MMSE scores than the other subtypes. Hippocampal-sparing AD patients showed a higher proportion of abnormal CSF phosphorylated tau levels as compared with limbic-predominant AD patients. The frequency of the A + T + N + pro le (68%) was signi cantly higher (p < 0.001) than the frequencies of A + T + N-(22%) and A + T-N-(10%) pro les in the ADNI cohort (Fig. 2). Interestingly, none of the ADNI patients had an A + T-N + pro le. In the KIDS cohort, the frequency of the A + T + N + pro le (55%) was also signi cantly higher (p < 0.001) as compared with the other ATN pro les. Interestingly, we observed a substantial proportion of A + T-N+ (15%) patients in the KIDS cohort. The A + T-N-pro le accounted for 30% of the KIDS patients, and the A + T + N-pro le included less than 1% of KIDS patients.

Association between AD subtypes, ATN pro les, and key demographic and clinical factors
Visual inspection of the correspondence between AD subtypes and ATN pro les showed that in ADNI, the A + T + N + was the most frequent group across all AD subtypes (Fig. 2). The A + T-N-group was present in every subtype but in minimal atrophy AD in ADNI. In KIDS, the frequency of A + T + N + was the lowest in typical and limbic-predominant AD (Fig. 2). In contrast, A + T-N + was equally distributed across all AD subtypes. The A + T + N-was only present in the typical AD subtype in KIDS.
These visual analyses were further supported by the FAMD models, also showing the correspondence between AD subtypes, ATN pro les, WMSA burden, age, sex, MMSE, cohort, and APOE genotype.
Dimension 1 explained 19% of the variance and was mainly driven by cohort and age (although WMSA burden, ATN, AD subtype, and sex also contributed statistically signi cantly to Dimension 1). Dimension 2 explained 12% of the variance and was driven by AD subtype (although MMSE, ATN, and cohort also contributed statistically signi cantly to Dimension 2). Dimension 3 explained 12% of the variance and was driven by AD subtype, MMSE, and ATN. Although ATN and AD subtype contributed to the three Dimensions, it was different categories within ATN and AD subtype that differently contributed to the Dimensions. To elaborate on this, Dimensions 2 and 3 were plotted against Dimension 1 (Figs. 3 and 4). Figure 3 shows that older patients have higher WMSA burden and tend to be from the ADNI cohort, clustering together. This cluster also showed a high frequency of A + T + N + and A + T + N-pro les, and a high frequency of patients with the typical AD subtype (Fig. 4). This cluster also includes a substantial proportion of patients with the hippocampalsparing AD subtype. However, the AD subtype factor is slightly oblique to Dimension 1, hence many hippocampal-sparing AD patients fall within a second cluster including younger patients with lower WMSA burden who tended to be from the KIDS cohort (Fig. 3). This second cluster showed a high frequency of A + T-N-and A + T-N + pro les (Fig. 4), and included most of the patients with limbicpredominant and minimal atrophy AD. However, as explained above, the AD subtype factor is slightly oblique to Dimension 1, so this second cluster included patients from typical and hippocampal-sparing AD subtypes as well.
Dimension 3 separates the AD subtypes and ATN pro les more clearly and shows the effect of MMSE. When Dimension 3 is plotted against Dimension 1, it can be observed that A + T + N + and A + T + N-have lower MMSE scores, independently of the cohort, WMSA burden, and age (Fig. 4). Limbic-predominant AD patients have higher MMSE scores, and typical and hippocampal-sparing AD patients have lower MMSE scores, while minimal atrophy AD does not completely align with MMSE scores (Fig. 4).
The complementary FAMD model adding APOE ε4 status was conducted in the subsample with available APOE data (N = 178). This model showed very similar results to the main FAMD model. Dimension 1 explained 19% of the variance and was mainly driven by cohort and age. Dimension 2 explained 12% of the variance and was driven by AD subtype. Dimension 3 explained 11% of the variance and was driven by ATN and APOE ε4 status. Within Dimension 3, APOE ε4 carriers showed a higher frequency of A + T-Nand A + T + N-pro les, and a tendency to include patients with typical AD.

Discussion
We investigated the association between AD subtypes and ATN classi cation scheme in two cohorts with different source characteristics. As hypothesised, the distribution of AD subtypes and ATN pro les differed between the research oriented cohort (i.e., ADNI) and the clinically oriented cohort (i.e., KIDS). In addition, we empirically tested the recent conceptual framework of biological subtypes of AD [5]. We applied a multivariate method for data analysis to investigate the association between AD subtypes, ATN pro les, and key demographic and clinical factors, including WMSA burden, age, sex, global cognition, and APOE genotype. To our knowledge, this study is the rst in investigating AD subtypes in combination with ATN pro les in patients with AD dementia.
The recent conceptual framework of biological subtypes of AD proposes two dimensions: severity and typicality [5]. The severity dimension corresponds to the "N" domain of the ATN scheme and includes typical and minimal AD as the two extremes of a continuum of neurodegeneration. Typical AD is in the severe end of the continuum, and minimal atrophy AD is in the other end. In our study, typical AD was the most frequent subtype in ADNI and KIDS. This result might seem unexpected since ADNI recruited mild to moderate AD patients, while typical AD would re ect full-blown AD at the highest degree of neurodegeneration. However, ADNI is a highly selective research cohort with strict inclusion criteria [12] that aimed to recruit the prototypical amnestic presentation of AD, which correlates with the typical AD subtype in neuropathological studies [32]. In addition, ADNI recruited patients with high education, which probably positively in uenced patients' cognitive reserve, possibly explaining why patients in ADNI have overt ATN and brain atrophy pro les, yet they are at mild to moderate clinical stages. On the other hand, the clinically oriented KIDS cohort is a naturalistic memory clinic sample that includes younger patients mainly at an early clinical stage with challenging differential diagnoses. This could explain the higher frequency of patients in the minimal atrophy AD subtype in KIDS.
The typicality dimension in the conceptual framework of biological subtypes of AD includes limbicpredominant AD on the one side, and hippocampal-sparing AD on the opposite side, both deviating from typical AD in the middle [5]. We found that limbic-predominant AD was slightly more frequent in KIDS, and hippocampal-sparing AD was slightly more frequent in ADNI. Based on previous studies [5], we hypothesised that these differences could be explained by demographic and clinical factors. To further test for this hypothesis, we investigated the association between AD subtype, ATN pro les, age, sex, cognitive status, and APOE genotype (discussed below).
A + T + N-and A + T + N + pro les were more frequent in ADNI than in KIDS. The current biological de nition of AD [39] postulates that A + is the rst pathological change, followed by T + and, eventually, N+. Further, A + and T + re ect AD pathology, while N + is unspeci c, with pathologies other than A and T contributing to neurodegeneration (N) as well. Hence, our nding of A + T + N-and A + T + N + being more frequent in ADNI than in KIDS could be related to the stricter selection criteria of ADNI, with a special interest on the amnestic form of AD. This interpretation is further supported by our nding of a high frequency of A + T-N + in KIDS. The N + category in the presence of a T-category suggests that the neurodegeneration in these patients is due to some pathology other than tau NFT, which suggests a mixed aetiology of clinical AD. As explained above, KIDS is a heterogeneous naturalistic memory clinic sample including young patients with challenging diagnoses, as re ected by the higher frequency of the A + T-N + pro le. Hence, the frequency of ATN pro les is highly dependent upon cohort, but not so much upon AD subtype.
CVD could be one of the non-AD pathologies contributing to N+. A previous study demonstrated that CVD contributes differently to AD subtypes [35]. Our current study provides novel data on the association between CVD, AD subtype, and ATN classi cation scheme. We found a higher WMSA burden in ADNI. This result may be unexpected since vascular risk factors (a predictor of WMSA) [40,41] are exclusion criteria in ADNI. However, previous studies showed that WMSA burden increases with older age [41,42], and ADNI patients are older than KIDS patients in our study, which could explain our nding of higher WMSA in ADNI. This nding aligns with the recent conceptual framework of biological subtypes of AD [5], i.e., older patients had higher WMSA burden, they more frequently had an A + T + N + pro le, and included a higher proportion of typical and limbic-predominant AD cases. Further, typical AD patients showed greater cognitive impairment as compared with limbic-predominant AD [5].
Our complementary FAMD model showed that APOE ε4 carriers tended to cluster together with patients with A + T-N-and A + T + N-pro les who belonged to the typical AD subtype. The association between the APOE ε4 genotype and amyloid-beta pathology (A+) is a well-established nding [43]. Further, previous studies showed that the frequency of APOE ε4 is higher in typical AD than in hippocampal-sparing AD [5]. Sex only marginally contributed to Dimension 1 in the main FAMD model. Although sex is also listed as one of the contributors to the emergence of AD subtypes [5], our current data suggest that the contribution of sex is less prominent than that of ATN pro les and other demographic and clinical factors. All in all, our ndings largely support the recent conceptual framework of biological subtypes of .
AD subtype and ATN classi cations are two popular approaches to disentangle disease heterogeneity in AD. An important nding in our study is that the correspondence between AD subtypes and ATN pro les is not absolute, suggesting that both approaches may capture complementary information. The FAMD model showed that AD subtype was the main driver of one of the dimensions (Dimension 2), while ATN always emerged as a secondary driver after AD subtype, MMSE, cohort, age, or WMSA burden (Dimensions 1, 2, and 3). The association of ATN with MMSE, age, and WMSA burden, as well as the ATN distribution observed in the highly selective homogeneous ADNI cohort suggest that the ATN classi cation scheme may by useful to assess disease staging. The capacity of AD subtype to drive a dimension by itself, partially independently of ATN and demographic and clinical factors, suggests that AD subtype classi cation may be less in uenced by disease staging. Whether AD subtypes re ect disease staging or truly distinct subtypes is an open discussion [2,4,32,44] that can only be answered in future longitudinal studies. The distinct subtypes hypothesis postulates that there are different pathophysiological pathways underlying clinical syndrome in AD [2,5]. Current data show that these pathways seem to rely on different forms of spread of pathology across the brain [5,32], leading to different patterns of brain atrophy in structural MRI [33]. An advantage of the AD subtype classi cation is the inclusion of the topographical dimension to the N category of ATN [35]. Whether AD subtype is a stronger approach to disentangle disease heterogeneity as compared with the ATN classi cation scheme must be con rmed in future studies.

Limitations
The current study has some limitations. We did not include A-individuals in the main analysis. Including A-individuals might increase the heterogeneity and shown slightly different associations between AD subtype, ATN pro les, and demographic and clinical factors. Further, the methods to assess WMSA were different in ADNI and KIDS. In the ADNI cohort we used an automatic segmentation based on white matter hypointensities while in the KIDS cohort we used visual ratings based on white matter hyperintensities. Although using different methods for WMSA could induce some noise in our analysis, we recently showed that both methods are strongly associated with each other [27]. Further, by classifying the output from both methods into high and low WMSA burden, we used a rougher measure that is less in uenced by differences between the two methods and has greater clinical applicability [42].
Finally, we lacked data for several factors listed in the recent conceptual framework for biological subtypes of AD [5]. Future studies should thus extend our current analysis by including measures of education or cognitive reserve, other markers of CVD, information about disease onset or disease duration, and data on speci c cognitive domains. Investigating the contribution of other comorbid brain pathologies such as Lewy body pathology or TDP-43 is challenging at present by the lack of reliable biomarkers for these two pathologies.

Conclusions
We conclude that the distribution of AD subtypes and ATN pro les depends on the source of the patients and it aligns with different demographic and clinical factors, depending on whether the cohort is more selective and homogeneous or more naturalistic and heterogeneous. Our ndings largely support the recent conceptual framework of biological subtypes of AD [5]. This framework postulates that the combination of risk factors, protective factors, and comorbid brain pathologies will determine belonging of AD patients to distinct biological subtypes of AD. Future studies should continue testing this framework with the goal of advancing our currently limited possibilities to realize precision medicine in clinical routine.

Declarations
Inc., Piramal Imaging, Servier, Synarc Inc. and Takeda   AD subtypes based on patterns of brain atrophy. Regional atrophy was measured with the MTA, PA and GCA-F visual rating scales based only on T1-weigthed images. In the three visual rating scales, a score of zero denotes no atrophy, whereas scores from one to three (PA and GCA-F) or four (MTA) indicate an increasing degree of atrophy. The typical AD subtype was de ned as abnormal MTA together with abnormal PA and/or abnormal GCA-F. The limbic-predominant subtype was de ned as abnormal MTA alone with normal PA and GCA-F. The hippocampal-sparing subtype included abnormal PA and/or abnormal GCA-F, but normal MTA. The minimal atrophy subtype was de ned as normal scores in MTA,    Factorial analysis of mixed data (FAMD): scatterplots for cohort and WMSA burden. Dots represent individual AD patients. In panels A, B, C and D, the x-axis represents Dimension 1. In panels A and B the yaxis represents Dimension 2. In panels C and D the y-axis represents Dimension 3. AD = Alzheimer's Disease; WMSA = white matter signal abnormalities; A + = CSF Aβ abnormal; T − = CSF p-tau normal; T + = CSF p-tau abnormal; N− = CSF t-tau normal; N + = t-tau abnormal. MMSE = mini-mental state exammination.

Figure 4
Factorial analysis of mixed data (FAMD): scatterplots for AD subtypes and ATN pro les. Dots represent individual AD patients. In panels A, B, C and D, the x-axis represents Dimension 1. In panels A and B the yaxis represents Dimension 2. In panels C and D the y-axis represents Dimension 3. AD = Alzheimer's Disease; WMSA = white matter signal abnormalities; A + = CSF Aβ abnormal; T − = CSF p-tau normal; T + = CSF p-tau abnormal; N− = CSF t-tau normal; N + = t-tau abnormal. MMSE = mini-mental state exammination.