> Cogniomics

Characterization and Diagnosis of Neurological and Neuropsychiatric Disorders


Time Line​

Ongoing since 2007


JD, Malin Björnsdotter, Srinivasan K, Francois B. Vialatte (ParisTech, EPSCI), Esteve Gallego (University of Vic), Jordi Casals (University of Vic), Moe Elgendi (now at University of Alberta), Jaeseung Jeong (KAIST), Andrzej Cichocki (RIKEN BSI, Japan)



We apply machine learning and signal processing to reveal brain patterns that characterize and predict neuropsychiatric and neurodegenerative disorders.  We primarily study electro-encephalography (EEG) and magnetic resonance imaging (MRI) data from patients and control subjects, but also other parameters such as behavioral measures and genetics. We focus on Alzheimer’s disease and mild cognitive impairment (MCI), but we also study autism spectrum disorder (ASD) and attention deficit disorder (ADD). We ask questions such as: Can we detect if an individual suffers from Alzheimer’s disease based on a brain scan? Can we predict if (and, if so, when) a person will convert from healthy aging to mild cognitive impairment and to Alzheimer’s disease? Can we find brain-based biomarkers that accurately predict whether a child has genetic risk of developing autism? 


We have examined the predictive value of EEG synchrony measures in distinguishing AD and MCI patients from age-matched control subjects. We have, for example, shown that Granger causality and stochastic event synchrony (SES) distinguishes MCI patients from controls at an accuracy of 83% (Figure 1; Dauwels et al., 2010).  More recently, we have shown that random subsampling is an efficient and accurate method for detecting brain regions where the gray matter density (measured by MRI) is reduced in MCI and AD patients (Figure 2; Björnsdotter et al., 2013), and that this approach can incorporate other modalities (such as a behavioral parameters; Björnsdotter et al., 2012). This work may eventually lead to brain-based, reliable, multimodal diagnostic tools and in the development of biomarkers for neuropsychiatric, neurodevelopmental and neurodegenerative disorders.

Figure 1: Granger causality and stochastic event synchrony measures (SES)derived from EEG data effectively distinguished individuals with mild cognitive impairment from control subjects (Panel A). Panel B shows the electrode placement map for the 21 channel EEG system, and panel C shows an example of an EEG signal in the time-frequency domain used to compute the stochastic event synchrony measures.

Figure 2. A) Results on classification of individuals with Alzheimer's disease (AD, top panel) and mild cognitive impairment (MCI, bottom panel) vs. healthy, age-matched controls, showing that hippocampus gray matter is compromised in these conditions. From Björnsdotter et al., 2013. B) Results on simulated data showing that our algorithm for fusing a clinical score with local brain activity patterns outperforms either measure on its own both in terms of classifying patients vs. controls (Peak classification) and identifying brain regions that are impaired in the disorder  (Mapping performance). From Björnsdotter et al., 2012.


C. Laske, H. R. Sohrabi, S. M. Frost, K. López-de-Ipiña, P. Garrard, M. Buscema, J. Dauwels, S. R. Soekadarl, S. Mueller, C. Linnemann, S. Bridenbaugh, Y. Kanagasingam, R. N. Martins, Sid E. O'Bryant, Innovative diagnostic tools for early detection of Alzheimer's disease, Alzheimer & Dementia, in press [ PDF ]

E. Gallego-Jutglà, J. Solé-Casals, F. Vialatte, J. Dauwels, A. Cichocki,  EEG based index for early diagnosis of Alzheimer's disease, Journal of Alzheimer's Disease 43:3, 2015. [ PDF ]

E. Gallego-Jutglà, J. Solé-Casals, F. Vialatte, M. Elgendi, A. Cichocki, J. Dauwels, A hybrid feature selection approach for the early diagnosis of Alzheimer's disease, Journal of Neural Engineering, in press. [ PDF ]

Björnsdotter, M, Malmgren, H. and Dauwels, J. (2013) Volume-of-interest subsampling for diagnostic brain mapping: application to the ADNI cohort, 19th Annual Meeting of the Organization for Human Brain Mapping.       

M. Björnsdotter, S. Rosenthal, D. Sona, J. Dauwels, Clustered subsampling for behaviorally informed diagnostic brain mapping, Fusion 2012 - 15th International Conference on Information Fusion, July 2012, Singapore [ PDF ].    
J. Dauwels,  Srinivasan K, Ramasubba Reddy M, T. Musha, F.Vialatte, C. Latchoumane, J. Jeong, A. Cichocki, Slowing and loss of complexity in Alzheimer's EEG: Two sides of the same coin?,  International Journal of Alzheimer's Disease, (invited paper), Volume 2011 (2011), Article ID 539621, doi:10.4061/2011/539621 [ PDF ].   

J. Dauwels, F. Vialatte, and A. Cichocki, Diagnosis of Alzheimer’s Disease from EEG Signals: Where Are We Standing?, Current Alzheimer's Research, 2010, 7(6):487-505 (invited paper) [ PDF ].

J. Dauwels, F. Vialatte, M. Vialatte, and A. Cichocki, A Comparative Study of Synchrony Measures for the Early Diagnosis of Alzheimer’s Disease Based on EEG, NeuroImage, 2010, 49:668-693 [ PDF ]. 

J. Dauwels, F. Vialatte, T.Weber, T.Musha and A. Cichocki, Quantifying Statistical Interdependence by Message Passing on Graphs PART II: Multi-Dimensional Point Processes, Neural Computation, 2009, 21(8):2203-2268 [ PDF ]. 

J. Dauwels, F. Vialatte, T. Rutkowski, and A. Cichocki, Measuring Neural Synchrony by Message Passing, In Advances in Neural Information Processing Systems 20, Edited by Platt J, Koller D, Singer Y, Roweis S, Cambridge, MA: MIT Press 2008:361-368 

[ PDF ]. 

E. Gallego-Jutglà, M. Elgendi, F. Vialatte, J. Solé-Casals, A. Cichocki, C. Latchoumane, J. Jeong, and J. Dauwels, Diagnosis of Alzheimer's Disease from EEG by Means of Synchrony Measures in Optimized Frequency Bands, EMBC 2012, San Diego, USA [ PDF ].

Localization of Seizure Focus from Interictal Intracranial EEG

Time Line​

Ongoing since 2008


JD, Jing Jin, Sydney Cash (MGH and Harvard Medical School), Andrew Cole (MGH and Harvard Medical School)



For approximately 30% of epilepsy patients, seizures are poorly controlled with medications alone. Those patients may be successfully treated by surgically removing the brain area(s) where the seizures originate; it is obviously crucial to accurately localize the seizure onset zone. To this end, one must often resort to semi-chronic invasive recordings of cortical activity, since non-invasive methods are frequently not conclusive. Currently neurologists rely heavily on seizures to determine the seizure focus. Since most patients have seizures infrequently, the invasive recordings must usually continue for days or weeks, until enough seizures are obtained; this procedure is costly, uncomfortable, and not without risk of side effects. 


Our long-term objective is to drastically shorten the hospitalization of epilepsy patients, from several weeks to a few days: We hope to be able to determine the seizure focus from short invasive recordings in the operating room, made before resection of the seizure focus. The goal of the proposed project is to explore the feasibility of this idea, building upon our promising preliminary results. 


Our hypothesis is that, even at rest, the seizure focus is characterized by seizure measures such as interictal spikes, high frequency oscillations (HFOs), slowing and synchrony. Our novelty is to exploit  combinations of all measures in a machine-learning framework, to localize the seizure foci. We have proposed methods for measuring synchrony and detecting spikes and HFOs respectively.

As a first step, we have applied signal processing techniques to invasive semi-chronic recordings between seizures, in order to extract signatures of the seizure focus, such as slowing and locally increased synchrony, spikes and HFOs. We have designed statistical decision algorithms that leverage those different signatures to determine the seizure focus in an automated fashion.

Figure: Semichronic EEG recording containing epileptic seizures

In future work, we will apply those algorithms to short invasive recordings made in the operating room. We will compare our results with the gold standard, determined by clinicians from invasive recordings during seizures.

The proposed procedure may have enormous impact on clinical practice of epilepsy in Singapore and elsewhere, and would substantially reduce treatment costs. Moreover, our novel automated approach to medical decision making is not only relevant for neurosurgery but many other medical disciplines.


J. Dauwels, E. Eskandar, A. Cole, D. Hoch, R. Zepeda, Sydney S. Cash. Graphical Models for Localization of the Seizure Focus from Interictal Intracranial EEG. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2011)  

[ PDF ]    

J. Dauwels, S. Cash, and E. Eskandar, Localization of Seizure Onset Area from Intracranial Non-Seizure EEG by Exploiting Locally Enhanced Synchrony, Proc. 31th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC09), in press. [

Jin, J., Dauwels, J., & Cash, S., Automated localization of the seizure focus using interictal intracranial EEG, Proc. 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC 2014), pp. 4439-4442).  [

Figure: (a) Signatures of epileptic EEG activity: spikes (top) ands HFO (bottom); (b) physical locations of surface electrodes, where the actual focus is red-color labeled by doctors from ictal EEG; (c) the distribution of a measure (e.g. spike rate) over all electrodes, where the corresponding ground truth is marked using back-color squares.

 Computational Modeling of Epileptic Seizures

Time Line​

Ongoing since 2012


P. N. Taylor, G. Baier (University of Manchester), S. S. Cash (MGH and Harvard Medical School), JD, J.J. Slotine (MIT), Yujiang Wang (University of Manchester)



Pathological brain rhythms such as those observed on the EEG of epileptic patients during a seizure are challenging to understand, and the seizures are difficult to treat clinically. Hypothesis driven, patient-specific treatments are desirable and computational modeling can represent an important step towards achieving this goal.


In our work we have developed models of epileptic brain activity which can reproduce many of the phenomena observed clinically in the spatial, temporal, and frequency domains. We have used our models to investigate hypotheses regarding stimulation protocols and to make explicit predictions with regard to heterogeneous brain activity.

Figure: Computational model of stimulus induced epileptic spike-wave discharges. (Left) Connectivity scheme of the model. (Right) Upper panel: time series of the cortical output containing a seizure, and time-frequency analysis using a Morlet wavelet transform. Lower panel: time series of a clinical recording of an epileptic seizure, and time-frequency analysis using a Morlet wavelet transform.


P. Taylor, Y. Wang, M. Goodfellow, J. Dauwels, F. Moeller, U. Stephani, G. Baier, A computational study of stimulus driven epileptic seizure abatement, PLOS One, Dec 2014, DOI: 10.1371/journal.pone.0114316.  [ PDF ].


P. N. Taylor, G. Baier, S. S. Cash, J. Dauwels, J.J. Slotine, Yujiang Wang, A model of stimulus induced epileptic spike-wave discharges, SCCI 2013, 2013 IEEE Symposium on Computational Intelligence, Cognitive Algorithms, Mind, and Brain, accepted. [ PDF ].   


Taylor P.N., Kaiser M., Dauwels J., Structural connectivity based whole brain modelling in epilepsy, Journal of Neuroscience Methods, 236, 51-57, 2014 (invited paper).  [ PDF ].  


Baier, G., Goodfellow, M., Taylor, P.N., Wang, Y., Garry, D.J., 2012, The importance of modelling epileptic seizure dynamics as spatio-temporal patterns, Frontiers in  Physiology.


Taylor, P.N., Goodfellow, M., Wang, Y., Baier G., 2012, Towards a large scale model of epileptic spike-wave discharges, Biological Cybernetics.


Goodfellow, M., Taylor, P.N., Wang, Y., Garry, D.J., Baier, G., 2012. Modelling the role of tissue heterogeneity in epileptic rhythms, European Journal of Neuroscience.


Wang, Y., Goodfellow, M., Taylor, P.N., Baier, G., 2012, A minimal phase space approach to model prototypic epileptic dynamics, Physical Review E.


Taylor, P.N., Baier, G., 2011. A spatially extended model for macroscopic spike- wave discharges, Journal of Computational Neuroscience.


Mathematical Approach to Controlling Epileptic Seizures

Time Line​

Ongoing since 2013


Justin Dauwels, Nishant Sinha (Nanyang Technological University), P. N. Taylor (Newcastle University, UK), S. S. Cash (MGH and Harvard Medical School), Justin Ruths (Singapore University of Technology and Design).



Nearly one-third of epilepsy patient continue having seizures despite treatment with anti-epileptic drugs. When the seizures consistently originate from a specific area of the brain (focal seizures), surgical removal of those cortical tissues might abate seizures. Predicting the likelihood of surgical success prior to performing surgery would be an important tool for neuro-surgeons toward achieving a successful surgery. As an alternative to surgery, stimulation of the brain at selected locations via electrical or other signals may lead to seizure abatement. To this end, it is crucial to determine suitable locations for stimulation, and to design appropriate stimuli for seizure abatement, while taking into account the heterogeneous dynamics of the human brain.


We developed computational models of focal seizures, based on patient data. Specifically, we infer the functional connectivity from the patient's ECoG and use it in computational models to replicate the dynamics of epileptic brain activities. Simulations were carried out in this model to determine brain areas that significantly reduce the seizure likelihood when resected. These regions were often found to be highly correlated with the clinically resected areas (see Figure 1).

We also develop computational models to simulate generalized seizures. We applied optimal control theory to design optimum stimuli to abate these simulated seizures. A moving window controller was derived, which monitors the system state in real-time. When a seizure is detected, the ensemble control is triggered which drives the state to a stable equilibrium, thereby stopping the seizure (see Figure 2).

Figure 1: The top panel on the left shows the simulated model dynamics & escape time for one of the channels when model parameters are inferred from the clinical interictal ECoG recording. The top right panel shows the optimal brain areas (in red) which should be resected to reduce seizure likelihood. These areas were found to be correlated with clinically determined seizure focus (in blue). The relative shift in the histogram of the bottom panel represents the reduction in seizure likelihood (or increase in escape time) in cases of no resection (in green), resection of simulated brain areas (in red) and random resection (in blue).

Figure 2: The top left panel shows a heterogeneous model of the brain with four cortical compartments and one thalamic compartment. The top right panel shows a moving window controller that detects and stops seizures as they occur. Boxplots represent the aggregate statistics for controlled and uncontrolled seizure durations over a wide range of kick parameters. The bottom panel on the left shows the uncontrolled seizures simulated from the heterogeneous model and the bottom panel on the right shows the controlled seizures when optimal stimuli (shown in the inset axis) were applied.


N. Sinha, P. N. Taylor, J. Dauwels, and J. Ruths, Development of optimal stimuli in a heterogeneous model of epileptic spike-wave oscillations, 2014 IEEE International Conference on Systems, Man, and Cybernetics (SMC2014), 2014, in press (invited paper).  [ PDF ]  

N. Sinha, J. Dauwels, Y. Wang, S.S. Cash, P. Taylor, An in Silico Approach for Pre-Surgical Evaluation of an Epileptic Cortex, 36th Annual International Conference of the IEEE Engineering In Medicine And Biology Society (EMBC 2014), 2014, in press. [ PDF ]   

J. Ruths, P. Taylor, J. Dauwels, Optimal Control of an Epileptic Neural Population Model, Proceedings of the International Federation of Automatic Control (IFAC), Cape Town, 2014.  

 Analysis of Paroxysmal Gamma Waves in Meditation EEG

Time Line​

Ongoing since 2011


JD, Jing Jin, Manuel Vazquez (Universidad Carlos III de Madrid), Francois B. Vialatte (ParisTech, EPSCI), Andrzej Cichocki (RIKEN BSI, Japan)



Meditation is a fascinating topic, yet has received limited attention in the neuroscience and signal processing community so far. Strong EEG activity has been observed in the left temporal lobe of meditators. Meditators exhibit more paroxysmal gamma waves (PGWs) in active regions of the brain. The knowledge of the PGW distribution may lead to a better understanding of the brain activity during meditation. However, comprehensive PGW marking is time consuming due to the high density, which motivates us to develop automated PGW detection methods. 


We are currently modeling traffic parameters as signals residing on a graph (spatial variable) instead of time series. We are developing different filtering techniques to analyze the behavior of traffic parameters such as speed, flow and travel time as spatial variables. Such analysis can potentially provide us with better mechanisms to model and limit traffic congestion in road networks.

We have also analyzed the synchrony between PGWs, revealing functional connectivity patterns in the brain during BhPr. Specifically the method of Stochastic Event Synchrony (SES) is applied to pairs of PGW sequences in order to assess their synchrony. From those pair-wise synchrony measures, large-scale functional connectivity patterns are inferred.

Strong synchrony can be observed in the temporal lobes for all subjects, in addition to long-range inter-hemispheric connections. Consistent connectivity patterns are present for exhalation periods, while those patterns are substantially less stationary for inhalation periods. Interestingly, the synchrony seems to increase gradually during the meditation session. Moreover, the distribution of synchrony values seems to depend on the level of expertise in practicing BhPr: the higher the expertise, the more concentrated the intensity values.

Figure 1: Examples of PGWs found in meditation EEG

Figure 2: Diagram of the proposed procedure (left) and PGWs from 2 different clusters via K-means (right).

Figure 3: Back projection of  PGW rate from (a) raw data; (b) single source; (c) multiple sources; (d) cluster 1 PGWs only and; (e) cluster 2 PGWs only.

Figure 4: Connectivity map with connection strength in color (left), and EEG (right) with extracted PGW sequences from a pair of (a) adjacent channels and (b) distant channels.

Figure 5: Power distribution and connectivity networks three subjects (B, I, and E) with three successive periods of (a) exhalations, and (b) inhalations. L is the total number of connections, and Navg is the average PGW count of each connection.


M.A. Vazquez, Jing Jin, J. Dauwels, F.B. Vialatte, Automated Detection of Paroxysmal Gamma Waves in Meditation EEG, ICASSP 2013. [ PDF ]

Jin Jing, J. Dauwels, F.B. Vialatte, A. Cichocki, Synchrony Analysis of Paroxysmal Gamma Waves in Meditation EEG, ICASSP 2014, May 4-9, 2014, Florence, Italy, in press. [ PDF ]

Flexible Near-Losless Multi-Channel EEG Compression using Wavelets and Matrix/Tensor Decompositions

Time Line​

Ongoing since 2010


JD, Srinivasan K, Andrzej Cichocki (RIKEN BSI, Japan), M Ramasubba Reddy (IIT Madras, India)



Electroencephalogram (EEG), the recording of electric potentials on the human scalp, is primarily used for disease diagnosis in clinical settings. Many other applications and studies also utilize EEG and hence the amount of recordings is increasing every day. For example, let us consider the intracranial EEG recordings made to evaluate the epilepsy brain surgery; a 320 micro- and macro-electrode array recording at 32kHz sampling rate with 18-bit resolution generates nearly 3 terabytes of data per day. Usually the person undergoing such surgery is monitored in NICU for several days and this amounts to a huge data. Storing, processing, analysis, and transmission of such EEG datasets is a humongous task. Compression of multichannel EEG is a straightforward solution for the above-mentioned problems, and a flexible compression system can be tuned appropriately to the application at hand.


First we formulated a two-dimensional representation for the single-channel EEG, and this formulation is extended further to multichannel EEGs in our subsequent works. The multichannel EEG is arranged in the form of 2D (matrix/image), 3D (volume/3-way tensor), and 4D (4-way tensor) to effectively exploit the redundancies (correlations) present in them. In the first stage of compression, we used wavelet transforms and matrix/tensor decompositions to exploit the redundancies from multichannel EEG; further, we deploy a coding stage to effectively represent the wavelet coefficients/decomposition components.  In the second stage of compression, the residual error obtained from the first stage is subjected to entropy coding after quantization for further compression. The quantization provides strict error control in the reconstructed signal and thus the reconstructed signal has sufficient accuracy required for most purposes. We noticed that tensor based compression schemes provide attractive compression performance and best error performance compared to other schemes. We are planning to tailor, evaluate, and deploy our compression algorithms in clinical settings in the near future.

Figure: Near-lossless reconstruction of EEG. The original signal is shown at the top. The error signal is superimposed on the reconstructions; the performance measures are mentioned at the right side are CR-Compression Ratio, PRD-Percent Root-Mean-Square Difference, PSNR-Peak Signal-to-Noise Ratio.

Figure: EEG compression performance of wavelet-based volumetric coding, subband specific arithmetic coding, and PARAFAC based coding. For the PARAFAC based approach, the average reconstruction error is about 6% for a compression rate of 10, while the worst-case reconstruction error is limited to 1% (not shown here) . The average reconstruction error is lower for subband specific arithmetic coding, however, the corresponding worst-casereconstruction error is substantially larger.


J. Dauwels, K. Srinivasan, M. Ramasubba R., and A. Cichocki, Near-lossless multi-channel EEG compression based on matrix and tensor decompositions, IEEE Transactions on Information Technology in Biomedicine, 2013.   [ PDF ]   

K. Srinivasan, J. Dauwels, M. Ramasubba R., Multichannel EEG compression: Wavelet-based image and volumetric coding approach, IEEE Transactions on Information Technology in Biomedicine, 2013.  [ PDF ]        

K. Srinivasan, Justin Dauwels, and M. Ramasubba Reddy, A two-dimensional approach for lossless EEG compression, Biomedical Signal Processing and Control, Volume 6, Issue 4, October 2011, Pages 387-394. 
 [ PDF ]

Nanyang Technological University


  • LinkedIn Social Icon
  • Facebook Social Icon
  • Twitter Social Icon
  • Google+ Social Icon

© 2016 Justin Dauwels & Dauwels Lab at NTU. All Rights Reserved.

Website Design by Vishnu Prasad Payyada and Priyanka Mehta