Alzheimer’s disease (AD) is a chronic neurodegenerative disorder that gradually impairs memory and cognitive performance, and eventually the ability to carry out the simplest tasks. AD is traditionally diagnosed based on clinical interviews and psychometric testing, which are recognized to be largely imperfect. Therefore, it is necessary to establish neuroimaging-based biomarkers to improve diagnostic precision. Resting state functional magnetic resonance imaging (rs-fMRI) has been used as a promising technique for automatic identification of AD from healthy controls (HC)^1,2,3. However, most of these classification methods are supervised, i.e. they require a priori clinical labels to guide classification. In this study, we adopted various unsupervised clustering methods using rs-fMRI connectivity to investigate hidden structures in the AD spectrum without relying on a priori clinical labels.Rs-fMRI data from Alzheimer’s disease neuroimaging initiative (ADNI) database (http://adni.loni.usc.edu/) was used in this study. The sample consisted of subjects with three progressive stages of cognitive impairment – early mild cognitive impairment (MCI, N=23), late MCI (N=29) and AD (N=13) – along with matched healthy controls (HC, N=31). Data was acquired in a Philips 3T MRI scanner with TR=3s, TE=30ms and slice thickness=3.3mm. Standard pre-processing steps were performed and mean fMRI time-series were obtained from 200 functionally homogenous brain regions (cc200 template⁴). Static and dynamic functional connectivity⁵ (SFC and DFC) were obtained between all pairs of brain regions. While SFC gives connectivity strength, variance of DFC (vDFC) gives the temporal variability of connectivity⁵, and has been shown to convey biologically relevant information⁶ which is distinct from static connectivity. Significant group differences were obtained in SFC (and vDFC) using ANOVA and only the top significant features (p<0.01) were further used in clustering analysis. The main idea of unsupervised clustering is to group objects in such a way that objects in the same group are more similar to each other than to those in other groups. In this study, three clustering methods were adopted, i.e., hierarchical clustering⁷, ordering points to identify the clustering structure (OPTICS)⁸ and density peak clustering (DPC)⁹. These methods were specifically chosen because they did not require a priori specification of the number of clusters. Since clustering accuracy is often lower in high dimensional feature spaces, feature selection methods were applied. A forward searching (FS) method was used by ranking features based on statistical significance and sequentially adding features for clustering. The optimal subset was then determined to be the one resulting in highest accuracy. However, this method did not guarantee a global optimum as statistical significance of individual features does not necessarily guarantee cluster separation when they are combined. To overcome this problem, a genetic algorithm (GA) was used in this study¹⁰. GA is a search heuristic method inspired by stochastic evolution theory. It starts from a set of randomly generated solutions and iteratively selects better solutions with larger objective values, which have been generated from crossover and mutation operations (Fig. 1). Clustering was applied on three types of features: (i) SFC and vDFC, (ii) clinical diagnostic labels, and (iii) phenotypic and genetic variables^11,12,13 (Table 1). The accuracy of the clustering and feature selection were assessed by computing the similarity¹⁴ of clustering between all three feature types. We hypothesized that clinical diagnosis must have high clustering similarity with neurobiological (connectivity) and phenotypic/genetic markers of disease.FS and GA methods were compared in terms of similarity obtained from different iterations (see Fig. 2). With FS, the curve oscillated dramatically, while with GA, a clearly step-wise convergence was observed. Also, GA led to larger peak similarity. OPTICS gave higher similarity compared to DPC and hierarchical methods (Table. 2) and the determined number of clusters was consistent with clinical diagnosis. Features selected by GA and OPTICS mainly included connections in all lobes of the brain and more specifically related to default mode network (DMN), e.g., medial temporal lobe (MTL), posterior cingulate cortex (PCC), precuneus, superior and inferior parietal gyri (Fig. 3). These findings are consistent with previous studies^1,2 showing alterations in DMN in MCI and AD. Selected phenotypic/genetic variables are shown in Table 1 and are consistent with known behavioral and genetic alterations on the AD spectrum^12,13,15. The similarity was highest between connectivity and phenotypic variables while clinical diagnosis had low similarity with both of them. This suggests that clinical diagnostic criteria for MCI and AD are not completely grounded in underlying neurobiological, neurobehavioral and genetic markers. Further, our framework using unsupervised clustering may be used to evaluate the fidelity of clinical diagnostic criteria in other brain-based disorders.