Biomedical Signal Processing and Control 61 (2020) 102056 Contents lists available at ScienceDirect Biomedical Signal Processing and Control jo ur nal homepage: www.elsevier.com/locate/bspc Multi metric functional connectivity analysis based on continuous hidden Markov model with application in early diagnosis of Alzheimer’s disease Fatemeh Jamaloo a , Mohammad Mikaeili a,∗ , Maryam Noroozian b a Department of Biomedical Engineering, Engineering Faculty, Shahed University, Tehran, Iran b Memory and Behavioral Neurology Department, Roozbeh Hospital, Tehran University of Medical Sciences, Tehran, Iran a r t i c l e i n f o Article history: Received 21 October 2019 Received in revised form 24 June 2020 Accepted 26 June 2020 Keywords: Alzheimer’s diagnosis Mild Cognitive Impairment (MCI) Functional Connectivity Continuous Observation HMM Resting-state EEG a b s t r a c t Functional connectivity (FC) is referred to as statistical dependencies between regions of interest. To investigate brain functional connectivity, there are many different connectivity metrics and researches show that the choice of the connectivity metric influences the results of the study and there is no golden rule of choosing the best connectivity metric. It is assumed that functional connectivity has a neural basis, and therefor is related to a variety of different neurological disorders like Alzheimer’s disease (AD) and Parkinson. AD is the most common neurodegenerative disorder. Cerebral cortex damage and synaptic plasticity disturbance in AD cause a decrease in functional connectivity. Mild Cognitive Impairment (MCI) is the first stage in AD progression and causes measurable decline in memory and cognitive abilities. In this study, a novel methodology is presented to combine several connectivity metrics with the goal of improving between-class discrimination. In the proposed method, temporal changes of multiple functional connectivity metrics are investigated along sliding windows by modeling it as the observation vector of a continuous observation hidden Markov model (HMM). The performance of the proposed method is evaluated using resting state eyes-closed EEG data from 7 MCI patients and 7 age-matched normal controls (NC). Group differences were investigated in five different frequency bands: delta, theta, alpha, beta, and gamma. Method analysis revealed that NC subjects and MCI patients are discriminated with accuracy of %95.9 ± 0.4 and %97.2 ± 0.5 over the alpha and gamma frequency bands respectively, using leave one subject out cross validation. These results indicate the proficiency of the connectivity metrics combination in distinguishing MCI from NC. © 2020 Elsevier Ltd. All rights reserved. 1. Introduction Functional Connectivity (FC) is defined as the temporal inter- relationship between spatially remote neurophysiological events (separated brain regions) regardless of the apparent physical con- nection between the regions [1]. While FC is typically analyzed by different statistical dependency methods among signals in cou- pled neuronal systems, there is a lack of consensus about the best method to describe neuronal couplings [2]. Recent studies have provided some evidences of temporal non- stationarity of brain functional connectivities [3–7]. The most commonly used strategies for examining dynamics in resting-state ∗ Corresponding author. E-mail address: Mikaili@Shahed.ac.ir (M. Mikaeili). functional connectivities are limited to describe their state tran- sitions over time which don’t characterize the intrinsic stochastic relationships among those temporally dynamic brain states quan- titatively [8]. Subsequent recent studies have shown that dynamic functional connectivity can be conceived as a multi-state process wherein, the connectivity patterns pass through multiple discrete states [5]. When a system lies in a specific state, it will not evolve randomly, but rather in a very constrained manner, toward a particular subse- quent configuration [9]. Further studies [4,8] have highlighted how the analysis of spatiotemporal patterns (i.e., temporal sequences of frames), which repeatedly occur over time, can capture the evo- lution of resting state networks (RSNs) better than conventional analysis of single spatial patterns [4,9]. Eavani [4] used HMM to train the dynamic functional connectivity networks using fMRI time series. Sourty [10] modeled the relationship between differ- https://doi.org/10.1016/j.bspc.2020.102056 1746-8094/© 2020 Elsevier Ltd. All rights reserved.