[1] 12w Cortical Atrophy Pattern-based Subtyping Predicts Prognosis of Amnestic MCI: An Individual-level Analysis

[1] 200w We categorized amnestic mild cognitive impairment (aMCI) patients based on cortical atrophy pattern and evaluated whether the prognosis differed across the subtypes.Furthermore, we developed a classifier that learns the cortical atrophy pattern and predicts subtypes at an individual-level. A total of 662 aMCI patients were clustered into three subtypes based on cortical atrophy pattern. Of these, 467 patients were followed-up for more than 12 months, and the median follow-up duration was 43 months. To predict individuallevel subtype, we employed a machine learning-based classifier with a 10-fold crossvalidation scheme. aMCI patients were clustered into three subtypes: medial-temporal (MT), minimal (Min), and parieto-temporal (PT) atrophy subtype. The PT subtype had higher prevalence of APOE4 carriers, amyloid PET positivity, and greater risk of dementia conversion compared to the Min subtype. The accuracy for binary classification was 89.3% (MT vs. Rest), 92.6% (PT vs. Rest), and 86.6% (Min vs. Rest). When we employed ensemble model of three binary classifiers, the accuracy for predicting the aMCI subtype at an individual-level was 89.6%. aMCI patients with the PT subtype were more likely to have underlying Alzheimer's disease pathology and showed the worst prognosis. Our classifier may be useful for predicting the prognosis of individual aMCI patients.

[1] 162w Amnestic mild cognitive impairment (aMCI), a transitional state between normal aging and Alzheimer's disease (AD) dementia (Petersen, 2004), is known to consist of a heterogeneous group of patients (Knopman et al., 2015;McGuinness et al., 2015;Nettiksimmons et al., 2014;Ota et al., 2016;Pusswald et al., 2013). Longitudinal data show that, among aMCI patients, some convert to AD dementia, while others remain in the MCI stage or even revert to normal cognition, suggesting that various etiologies underlie cognitive impairment (Knopman et al., 2015;Vos et al., 2015). In clinical practice, early intervention is needed in individuals who are likely to develop neurodegenerative processes, but timely diagnosis is challenging. Although PET and cerebrospinal fluid (CSF) analyses enable the detection of AD biomarkers (amyloid or tau) (Shim and Morris, 2011), these methods are not yet widely available due to their high cost and invasiveness. Patterns of cortical atrophy, which precedes cognitive impairment reflect the underlying pathology and can be used as a reliable biomarker to predict patient prognosis.

[2] 108w AD-like cortical atrophy in aMCI can be a sign of underlying AD processes (Dong et al., 2016b;Zhang et al., 2012). However, only a few studies have examined anatomical subtypes of aMCI and their clinical implications (Dong et al., 2016b;Nettiksimmons et al., 2014). Using Alzheimer's Disease Neuroimaging Initiative (ADNI) data, Dong et al. classified aMCI and AD dementia patients into four anatomical subtypes and found that each subtype had distinct clinical features (Dong et al., 2016b). Nettiksimmons et al. also used ADNI data and identified four clusters of aMCI patients based on brain structural volume and CSF AD biomarkers (Nettiksimmons et al., 2014). These previous studies performed group analysis

[1] 132w Patients were clustered into three subtypes (Figure 1a). When we repeated the subtyping procedure 1,000 times for the major voting scheme, the labels were identified with a high consistency of 92.9% (Figure 1b). We called the first subtype the medial temporal atrophy (MT) subtype (n=208, 31.4%) because atrophy was mainly localized to bilateral medial temporal areas, especially in the parahippocampal, uncus, and fusiform gyri, compared to the NCs. The second type was the minimal atrophy (Min) subtype (n=258, 39.0%) in which patients did not show significant cortical thinning compared to NCs. The third type was the parieto-temporal atrophy (PT) subtype (n=196, 29.6%), which showed atrophy predominantly located in the parietal areas (superior and inferior parietal lobules and precuneus) and temporal areas (superior, middle, and inferior temporal gyri) compared to NCs (Figure 2).

[2] 130w We further analyzed 467/662 (70.5%) aMCI patients who were followed up for more than 12 months. Disease duration and follow-up duration did not differ across the three subtypes (Table 1). Among 467 aMCI patients, 171 (36.7%) patients developed dementia and 66 (14.1%) patients reverted to normal cognition within median follow-up of 43 months. Of 171 patients who developed dementia, 162 patients developed AD dementia and 9 patients developed non-AD dementia, which includes 3 vascular dementia, 1 normal pressure hydrocephalus, 3 progressive supranuclear palsy syndrome, 1 semantic variant primary progressive aphasia, and 1 dementia with Lewy bodies (Table 1). Patients with the PT subtype were at higher risk of dementia conversion compared to the Min subtype, which was significant even after controlling for age and education level (Table 2, Figure 1).

[3] 55w Our proposed method was highly reproducible and provided consistent results on the mean value for 10 subsets with 10% random removal. Among the various subtyping methods, our method (Louvain method) with correlation coefficients between cortical atrophy patterns had the largest modularity and highest reproducibility (92.94% reproducibility) and outperformed the conventional hierarchical clustering method (81.81% reproducibility)

[1] 88w In this large aMCI cohort, we identified three anatomical subtypes of aMCI: the MT subtype (31.4%), Min subtype (39.0%), and PT subtype (29.6%). Among the three subtypes, patients with the PT subtype were more likely to have underlying AD pathology and showed the worst prognosis. The overall accuracy for predicting the aMCI subtype at an individual level was 89.6%. These findings therefore emphasize the potential value of clustering patients based on brain atrophy pattern, which could contribute to recruitment of more homogenous patient populations for aMCI clinical trials.

[2] 101w Although a few previous studies investigated anatomical subtypes of MCI, our study has several advantages over the previous two studies (Dong et al., 2016b;Nettiksimmons et al., 2014). First, we applied a reliable cortical atrophy pattern-based clustering method. Our clustering method (the Louvain method with correlation) has advantages over conventional methods including its hierarchical clustering, increased robustness against sampling bias, and the ability to distinguish subjects based on cortical atrophy pattern (which represents true anatomical subtype) rather than overall atrophy severity (which is largely influenced by disease stage) (Park et al., 2017). Second, one of the previous studies clustered late aMCI and

[1] 52w We consecutively collected data from 740 aMCI patients from the Memory Clinic at Samsung Medical Center from July 2007 to December 2012 who underwent detailed neuropsychological testing and brain MRI including 3-dimensional T1 images. All patients were selected based on the following inclusion and exclusion criteria and were > 45 years old.

[2] 27w The patients met the Petersen's clinical criteria for MCI (Petersen et al., 1999) with the following modifications: (1) subjective memory problems reported by the patient or caregiver,

[3] 69w (2) normal activities of daily living (ADL) as judged by an interview with a clinician and the Seoul-Instrumental ADL test (with a score < 8) (Ku HM, 2004) M A N U S C R I P T A C C E P T E D ACCEPTED MANUSCRIPT Mental Disorders (Fourth Edition) criteria for psychotic or mood disorders, such as schizophrenia or major depressive disorder (Frances et al., 2000).

[4] 21w Of the 740 patients who met the above mentioned criteria, 78 were excluded because of segmentation errors during cortical thickness analysis.

[5] 85w Therefore, the final sample consisted of 662 aMCI patients. 320 normal controls (NCs) were selected from individuals who visited the Health Promotion Center of the Samsung Medical Center from September 2008 to December 2012. Inclusion criteria for NCs were as follows: (1) no complaint of subjective memory problems; (2) normal cognitive function, defined by a Mini Mental State Examination (MMSE) score above the 16th percentile for age-and education-matched norms; (3) no history of neurologic or psychiatric disorders; and (4) no structural abnormalities on brain MRI.

[6] 13w This study was approved by the Institutional Review Board of Samsung Medical Center.

[7] 34w All aMCI patients underwent neuropsychological tests using a standardized neuropsychological battery called the Seoul Neuropsychological Screening Battery (SNSB) (Ahn et al., 2010;Kang and Na, 2003), the details of which are described in Appendix e-1.

[8] 149w We analyzed the reproducibility of the proposed clustering method, which captures how a method provides consistent results. We measured the fraction of consistently assigned subjects to each subtype on average. Our dataset was divided into 10 subsets randomly, and we repeatedly performed our method 10 times excluding one subset in turn. Next, we computed the average fraction over 10 runs. The subtype label was considered to be reproducible if the subtype label rendered from each subset matched the subtype identified using all of the subjects. We further compared our method (the Louvain method with correlation) with the hierarchical clustering-based method (Noh et al., 2014) and performed the same experiment using the two different methods. For hierarchical clustering, we followed the method suggested by Noh et al. (Noh et al., 2014). The only difference was that we used our estimated cortical atrophy instead of raw cortical thickness for fair comparison.

[9] 12w We also performed the experiment using two different similarity metrics: the correlation

[10] 22w Cortical atrophy pattern-based subtyping of aMCI coefficient and Euclidean distance for investigating the effects of the correlation coefficients as a similarity measure.

[11] 61w To predict individual-level subtype, we employed a machine learning-based classification algorithm. Since we have three subtypes to predict at an individual level, we developed an ensemble model of three binary classifiers. Specifically, each binary classifier was constructed for discrimination of each subtype versus the other subtypes. Then, the individual subject's subtype was predicted by combining the results of three binary classifiers.

[12] 86w Our binary classification is based on principal component analysis (PCA) and linear discriminant analysis (LDA) with nested cross validation (Groth et al., 2013;Liu and Wechsler, 2000;Yu and Yang, 2001). PCA is a statistical dimension reduction method achieved by computing principal components that have the largest variance in the projection of data distribution by linear transformation. LDA finds the axis where the discrepancy within the class is maximized while the variance among the class is minimized. LDA is a commonly used linear classifier for binary classification problems.

[13] 39w In this study, we employed nested cross-validation for a performance validation scheme and hyperparameter tuning by dividing the dataset with 10% random removal. We further We finally determined a subtype label using a major-voting scheme with all 11 classifiers.

[14] 94w Second, an outer loop was used for evaluating our subtype prediction model with the unseen test set. During the outer loop, we repeated de novo group-level subtyping using only the training dataset. We employed the nested cross-validation scheme in order to avoid so called "circular performance evaluation" in the subtype prediction. That is, the ground truth label for each test subject was determined based on the whole dataset in the outer loop, while that of the training dataset was derived only from the training dataset in the inner loop without seeing the test dataset.

[15] 34w To evaluate whether age, gender, education level, and vascular risk factors differed across the anatomical subtypes, we performed analysis of variance (ANOVA) or Chi square followed by Bonferroni's post hoc analysis to compare groups.

[16] 22w To assess whether dementia risk differed across anatomical subtypes, we applied Cox's proportional hazards model after controlling for age and education level.

[17] 27w We used MatLab (Version 2014b, Mathworks, Natick, MA, USA) for MRI preprocessing and subtyping (http://bia.korea.ac.kr/software/AD_subtyping/). The SurfStat toolbox was used to visualize the cortical atrophy pattern (http://www.math.mcgill.ca/keith/surfstat).

[1] 18w and thus have limited value to individual patients because their findings cannot be directly translated to clinical practice.

[2] 63w Herein, we clustered aMCI patients based on similarities in cortical atrophy patterns. We used the graph-theoretical clustering method (Louvain method), which is robust against sampling bias (Park et al., 2017). We first compared clinical phenotypes between the anatomical subtypes and identified subtypes that have poor prognosis. More importantly, we proposed an individual subject classification method that classifies aMCI patients using cortical atrophy pattern.

[3] 91w We used 18 F-florbetaben PET (48 patients) or 11 C-Pittsburgh compound B PET (36 patients) to detect amyloid in the brain. We defined florbetaben PET as positive when visual assessment scored 2 or 3 on the brain Aß plaque load (BAPL) scoring system (Barthel et al., 2011). We defined Pittsburgh compound B PET as positive when the global uptake ratio (using cerebellar gray matter as the reference region) was more than two standard deviations from the mean of the normal controls (standardized uptake value ratio ≥ 1.5) (Lee et al., 2011).

[4] 20w Presence or absence of vascular risk factors such as diabetes, hypertension, and hyperlipidemia was defined by self or caregiver's report.

[5] 66w Genomic DNA was extracted from peripheral blood leukocytes using the Wizard Genomic DNA Purification kit following the manufacturer's instructions (Promega, Madison, WI). Two single nucleotide polymorphisms (SNP; rs429358 for codon 112 and rs7412 for codon 158) in the APOE gene were genotyped using TaqMan SNP Genotyping Assays (Applied Biosystems, Foster City, CA) on a 7500 Fast Real-Time PCR System (Applied Biosystems) according to the manufacturer's instructions.

[6] 14w Of the 662 aMCI patients, 467 (70.5%) were followed-up for more than 12 months.

[7] 81w The median follow-up duration was 43 months. The point of dementia conversion was determined by neurologists based on clinical interviews and neuropsychological tests M A N U S C R I P T A C C E P T E D ACCEPTED MANUSCRIPT Cortical atrophy pattern-based subtyping of aMCI including the instrumental ADL scale. For patients who did not undergo detailed neuropsychological tests (n=78), neurologists determined dementia conversion based on clinical interviews and the Geriatric Deterioration Scale (Choi et al., 2002).

[8] 75w An Achieva 3.0-Tesla MRI scanner (Philips, Best, The Netherlands) was used to acquire 3-dimensional T1 turbo field echo (TFE) MRI data from all participants using the following imaging parameters: sagittal slice thickness, 1.0 mm; over contiguous slices with 50% overlap; no gap; repetition time, 9.9 ms; echo time, 4.6 ms; flip angle, 8 degrees; and matrix size of 240 x 240 pixels reconstructed to 480 x 480 over a field of view of 240 mm.

[9] 130w We extracted surfaces of grey and white matter from T1-weighted images using FreeSurfer v5.1.0. (Dale et al., 1999;Dale and Sereno, 1993). We manually checked the quality of the extracted cortical surfaces and corrected them if necessary. The surface meshes were then resampled using 40,962 vertices for each hemisphere. We also employed the Laplace-Beltrami operator for diffusion smoothing of cortical thickness data with our in-house software, as in our previous studies (Cho et al., 2012;Qiu et al., 2006). We adopted our previous clustering method proposed for subtyping very mild AD patients (Park et al., 2017). This subtyping method creates clusters based on the similarity of cortical atrophy pattern with a graph-theoretical technique. First, we estimated the cortical atrophy patterns of aMCI patients using the cortical thickness data of the NC subjects.

[10] 22w Specifically, the z-score at each vertex of the cortical surface was calculated from the distribution of cortical thickness in the NC group:

[11] 143w Cortical atrophy pattern-based subtyping of aMCI is the z-score of the i-th vertex of the j-th aMCI patient; i=1,2, …, 81924, is the cortical thickness value of the j-th aMCI patient at the i-th vertex; and and are the mean and standard deviation of cortical thickness at the i-th vertex of the NC group, respectively. We then constructed a similarity matrix based on the correlation coefficients for cortical atrophy patterns between any pair of aMCI patients. Finally, we extracted clusters from the similarity matrix by exploiting the Louvain method (Figure 1). The Louvain method determined the modular organization by maximizing the modularity value Q with denser intra-modular connections and sparser inter-modular connections. While the Louvain method is accurate and efficient for large networks, the clustering results vary slightly when analyzed multiple times because the Louvain method finds clusters with a greedy optimization method.

[12] 74w We resolved this problem by using a major voting scheme. Each subject was labeled by the clustering result, which is the most frequently chosen in N repetition results. In addition, the Louvain method is able to manage the number of clusters by controlling the resolution parameter, γ. Similar to our previous study (Murray et al., 2011a;Park, 2017), we controlled this parameter (γ=0.9) and repeated the subtyping procedure 1000 times for the major voting scheme.

[13] 46w It is widely accepted that intra-cluster distance and inter-cluster distance are important measures of the quality of a specific clustering. The intra-cluster distance decreases as homogeneity within clusters increases, while the inter-cluster distance increases as homogeneity between different clusters decreases. A good cluster should minimize intra-

[14] 65w Cortical atrophy pattern-based subtyping of aMCI Kim et al 10 10 cluster distance and also maximize inter-cluster distance at the same time. The gamma value in the Lurvain's method controls the relative weight for intra-cluster and inter-cluster variance to determine the clusters. We performed cluster analysis using a range of gamma values (from 0.7 to 1.0), and examined the clinical implications of the resulting clusters.

[15] 52w Since conversion to AD dementia is the most important clinical implication in our problem setting, we used whether the patient converted to AD dementia as a metric for measuring inter-/intra-cluster distance. We selected three clusters as the optimal clustering number as it had the largest ratio of the inter-/intra-cluster distance (Table e-1).

[16] 11w Cortical atrophy pattern-based subtyping of aMCI Kim et al 13 13

[17] 14w There were no differences in age, gender, or education level across the three subtypes.

[18] 63w However, compared to the Min subtype, patients with the PT subtype were more likely to be APOE e4 carriers (p<0.001). Furthermore, patients with the PT subtype were more often positive for amyloid PET, with an amyloid PET positivity of 72.4% as opposed to 42.4% in the Min subtype (p=0.017) (Table 1). Baseline cognitive function did not differ across the three subtypes (Table e-2).

[19] 16w We achieved a high classification accuracy in our large aMCI cohort and summarized the performance in

[20] 94w Table e-4. First, we evaluated binary classification performance and M A N U S C R I P T A C C E P T E D ACCEPTED MANUSCRIPT Cortical atrophy pattern-based subtyping of aMCI achieved accuracy values of 89.3% (MT vs. Rest), 92.6% (PT vs. Rest), and 86.6% (Min vs. Rest). Second, we predicted individual patient subtypes according to the most frequently assigned in a set of candidate binary classifiers. When we employed ensemble model of three binary classifers the overall accuracy for predicting the aMCI subtype at an individual level was 89.6%.

[21] 10w Cortical atrophy pattern-based subtyping of aMCI Kim et al 16

[22] 38w AD dementia based on voxel-based morphometry and the CHIMERA method using 80 ROIs (Dong et al., 2016a;Dong et al., 2016b), whereas we applied surface-based morphometry, which is more sensitive for detecting cortical atrophy and we used approximately 80,000

[23] 97w vertices, an approach more suitable for high-dimensional data (Park et al., 2017). Third, our clustering method can be easily applied in clinical practice since it is based only on cortical thickness, which can be acquired through brain MRI and does not require incorporation of other biomarkers such as CSF amyloid/tau or molecular PET imaging (Nettiksimmons et al., 2014). Finally, none of the previous methods proposed an individual-level classification. Our classifier achieved a high accuracy of 89.6%, and we believe our method might serve as a useful means to predict the subtype and prognosis of individual aMCI patients.

[24] 182w Patients with the Min subtype did not show significant atrophy compared to NCs, had a relatively stable clinical course, and had the lowest frequency of amyloid PET positivity and In contrast, the PT subtype with a typical AD-like atrophy pattern proved to be a malignant type of aMCI. These patients had the highest frequency of amyloid PET positivity, APOE4 carriers, and showed the highest dementia conversion rate. This result is in line with previous studies showing that, among AD dementia patients, those with parietal-dominant atrophy showed rapid cognitive decline (Na et al., 2016). A previous study that clustered anatomical subtypes of aMCI and AD dementia together showed similar results in that the subtype with parieto-temporal atrophy showed classical AD characteristics with the fastest clinical progression (Dong et al., 2016b). Interestingly, our PT subtype patients were not significantly different in age from the other subtypes. Previous studies on anatomical or pathological AD dementia subtypes showed that patients with parietal-dominant atrophy exhibited younger symptom-onset age (Noh et al., 2014). Our results suggest that, in aMCI patients, parieto-temporal atrophy predicts worse prognosis regardless of age.

[25] 138w Our third type of aMCI, the MT subtype, consisted of individuals with localized atrophy in the medial temporal areas. This subtype is consistently reported in most aMCI cluster studies (Dong et al., 2016b;Nettiksimmons et al., 2014), even in AD dementia (Murray et al., 2011b;Noh et al., 2014) or SMI (Jung et al., 2016) cluster studies, and has been referred to as 'limbic-predominant' (Murray et al., 2011b) or 'medial-temporaldominant' atrophy (Noh et al., 2014). Previous studies of AD dementia consistently showed that patients with focal medial temporal lobe dysfunction have a relatively slow rate of progression (Butters et al., 1996;Na et al., 2016). Thus, one possibility is that MT subtype might represent the earliest stage of this process. The underlying pathology of our MT subtype patients might be AD (Murray et al., 2011b). However, we also suspect that these

[26] 22w Cortical atrophy pattern-based subtyping of aMCI patients might have hippocampal sclerosis or primary age-related tauopathy (Jellinger et al., 2015;Landau et al., 2016).

[27] 63w Although aMCI could be classified into distinct anatomical subtypes, we could not find differences in demographics or neuropsychological test results across the anatomical subtypes of aMCI. Cognitive scores might not capture regional differences in atrophy patterns and may lack the ability to detect heterogeneous atrophy patterns. Our results therefore suggest that anatomical classification can provide a more sensitive approach for defining aMCI subtypes.

[28] 48w Furthermore, although APOE genotype, amyloid PET, or CSF level of amyloid and tau may predict prognosis of aMCI patients, our classification algorithm may be more widely used because it only requires MRI data. These findings further emphasize the potential value of clustering patients based on brain atrophy pattern.

[29] 92w To the best of our knowledge, this is the first study to identify the anatomical subtypes of aMCI on an individual basis. The proposed method showed high accuracy for classification of aMCI subtypes (89.6%), which suggests that individual-level classification of aMCI subtypes based on distinct topographic patterns of cortical atrophy can be achieved by a simple but powerful method using LDA and PCA. More specifically, our binary classification models performed very well for distinguishing the PT subtype versus other subtypes (92.6%). This suggests that PT subtype has a unique characteristic atrophy pattern.

[30] 217w Our individual-level classifier has several advantages. First, it is an efficient method of analysis in the context of data accumulation. Previously, whenever researchers investigated the characteristics of each aMCI subtype, it was inevitable to perform a de novo cluster analysis at group-level. This is a crucial drawback in accumulating knowledge of each We acknowledge the limitations of this study. First, since we retrospectively collected data, the follow-up duration varied. However, there were no differences in follow-up duration across anatomical subtypes. Second, the sample was derived from a Memory Clinic, which may limit the accuracy of subtype classification in a more heterogeneous sample with greater vascular burden. Third, as the three subtypes might reflect different stages of the disease, further studies with longer follow-up duration is necessary. Finally, only a small number of patients underwent amyloid PET, and our data lacked tau biomarkers to confirm the underlying pathology. Table 2. Hazard ratios of dementia conversion according to anatomical subtype Model 1 Model 2 Model 3 HR (95% CI) p HR (95% CI) p HR (95% CI) p Anatomical subtype Minimal subtype REF REF REF Medial temporal subtype 1.08 (0.74-1.59) 0.695 1.10 (0.75-1.61) 0.630 1.08 (0.74-1.59) 0.687 Parieto-temporal subtype 1.51 (1.06-2.15) 0.024 1.47 (1.03-2.10) 0.033 1.45 (1.01-2.07) 0.042 Age 1.02 (1.00-1.04) 0.043 1.02 (1.00-1.04) 0.041 Education 1.02 (0.99-1.05) 0.100