University of Florida Has Been Hosting a Series of Conferences on Biocomputing, Biomedicine

DIMACS* Workshop on Data Mining, Systems Analysis and Optimization in Neuroscience

Organizers:

W. Art Chaovalitwongse, Rutgers University

Leonidas D. Iasemidis, Arizona State University

Panos Pardalos, University of Florida

University of Florida, Gainesville, FL

15-17 February 2006

Report Author: Onur Seref (University of Florida).

Date of Report: June, 2006

*DIMACS was founded as a National Science Foundation Science and Technology Center. It is a joint project of Rutgers University, Princeton University, AT&T Labs-Research, Bell Labs, NEC Laboratories America, and Telcordia Technologies, with affiliated partners Avaya Labs, Georgia Institute of Technology, HP Labs, IBM Research, Microsoft Research, Rensselaer Polytechnic Institute, and Steven Institute of Technology.

1- Introduction

University of Florida has been hosting a series of conferences on biocomputing, biomedicine and neuroscience with a focus on quantitative methods for more than five years. The last conference , titled “DIMACS Workshop on Data Mining, Systems Analysis, and Optimization in Neuroscience”, was held between February 15-17, 2006 in University of Florida. It was jointly sponsored by the Center for Discrete Mathematics and Theoretical Computer Science (DIMACS), the Biological, Mathematical, and Physical Sciences Interfaces Institute for Quantitative Biology (BioMaPS), and the Rutgers Center for Molecular Biophysics and Biophysical Chemistry (MB Center) under the auspices of the DIMACS/BioMaPS/MB Center Special Focus on Information Processing in Biology. Also Center for Applied Optimization, College of Engineering, Genetics Institute, Office of the Vice President for Research were among local sponsors from University of Florida.

A large variety of leading research topics in neuroscience were presented. Plenary talks on fundamental aspects of the scope of the conference addressed subjects such as the organizational structure of the transition from microscopic to macroscopic level [1]; probing brain functions at different spatial and temporal scales through single trial experiments [2]; and the wiring in the brain as optimal design problems solved by evolution [3]. Each plenary session ended with inspiring questions and an extended discussion. Although the talks followed a multi-disciplinary structure in general, they can be grouped under the main topics of optimization, systems analysis and data mining parallel to the main themes of the conference. Under optimization, the main emphasis was on classification and clustering methods and statistical approaches.

2- Optimization

The data regarding brain functions have a highly non-linear and high-dimensional nature, and thus, often require efficient and sophisticated optimization techniques for a translation from electrical signals into meaningful interpretation of the underlying mechanisms. Classification and clustering methods offer a powerful way to enable such translations. These methods involve optimization models at their core, which deliver the desired translation function in the most efficient way with the most precise results achievable.

Talks that revolved around classification introduced novel approaches with significant results. Improvements over optimization based solutions to discrimination problems, which involve mixed integer programming formulations for multiple hyperplane classification, were proposed [4], together with extensions on decision trees [5] and comparative computational results [6]. Single trial based experiments also provide a natural ground for classification. Applications of kernel based methods were presented for understanding the integration of visual and sensory-motor cortical areas through intracranial single trial neural recordings from monkeys [7]. One-class machine learning techniques were demonstrated on a similar set of experiments involving visual and motor areas of the human brain, using fMRI [8]. New Bayesian network based classification techniques for high dimensional data were introduced. The first of the two talks on this subject featured a score-based structure inference to study language acquisition in songbirds [9], while the second talk used a graphical model to combine a tabu search metaheuristic with a Markov blanket method [10].

Being very similar to classification algorithms in practice, novel clustering methods also found important application areas in mining brain imaging problems. A new fuzzy hyperplane clustering algorithm that uses a union of hyper-planes was shown to have direct applications in sparse representation problems, and successful results on fMRI data analysis were presented [11]. An image-pixel clustering method, which uses level set methods in a hierarchical structure, was introduced together with the results of a straightforward parallel implementation for brain imaging problems [12]. A quadratic optimization model was proposed to study the brain clustering problem, which has a potential application in differentiating the normal and pre-seizure states of an epileptic brain through analysis of EEG data [13].

3- Systems Analysis

The brain is a very complex system with a high level of uncertainty. Therefore, one of the most convenient ways of analysis is building statistical models for the variability of the brain functions over time and their distribution over different cortexes. An alternative approach to analyze such functions is through simulation using artificial neural networks. Some other network based approaches such as coherence methods or finding cliques in the brain also provide means to understand the brain on different scales as a highly connected system.

Statistical approaches, in general, comprise a large majority of quantitative studies on neural data, as they constituted the core of many talks during the conference. A generative data model that explains the variability of evoked responses was introduced in a classical statistical framework [14]. An extension followed this talk with more focus on current source density analysis with results on simulated data, as well as monkeys performing an intermodal selective attention task [15]. Different adaptations of stationary analytics methods were applied on a decision-making task in rats to understand nonstationary neural activity [16]. Spatio-Temporal changes are important indicators in predicting epileptic seizures. A statistical technique that involves a mantel test statistic on a non-linear synchronization measure was introduced to detect such changes before the seizure happens, up to more than an hour in advance [17].

Intuitively one of the best ways to model the brain is as a complex network, or to use network based models to solve problems regarding neural data, in different scales of connectivity in the brain. Using artificial neural networks is a well established data mining technique, however may be slow for large datasets because of slow convergence. Combining Monte Carlo with artificial neural network is capable of overcoming this problem with comparable generalization results [18]. Exploratory simulation tools that use an integrate and fire model were shown to be very efficient, even with a massive number of neurons, in understanding the influence of anatomy on information flow [19].

The existence of networks on varying scales of complexity and magnitude is a well studied, and yet mostly unexplored research area. One such network is the brain’s language processing system. Although the basic large scale structure is well known in individuals without language deficits, language processing may be fundamentally different in individuals with dyslexia. A comprehensive study using magnetoencephalography was presented that shows a clear distinction in the network structure of a normal and a dyslexic brain [20], with further verification using a power and a coherence study [21]. The highly connected structure of the brain sometimes causes more serious problems through synchronization of large scale networks such as in the case of epileptic seizures. This underlying network structure may be revealed by finding maximum cliques that shows high pair-wise spectral similarity, and a brain similarity network can be formed to discover the pattern of epileptic seizures efficiently [22].

4- Data Mining

As the rising level of functional brain imaging and data acquisition technology enable high-speed and high-resolution data to be available to researchers, major challenges result from a data mining perspective. The talks under the data mining topic involved important application areas such as epilepsy detection and control systems and visualization techniques that involve high resolution imaging, spatiotemporal tracking and monitoring plastic deformation in cortical tissues.

Epilepsy and related problems are one of the central research areas in neuroscience. Aside from the visiting speakers, there was a significant local contribution from the speakers affiliated with McKnight Brain Institute, a part of the Shands research hospital at University of Florida. The majority of the talks evolved around quantitative techniques in analyzing the underlying mechanisms of epilepsy and prediction of seizures before they happen. One of the important results pointed out that seizures serve as dynamical resetting mechanisms of the brain, which increased the knowledge base on epileptogenesis, seizure intervention and control, and even intermittent spatiotemporal state transitions in other complex systems [23]. Another notable approach was through simulation of an epileptic brain by modeling it as an interconnected network of nonlinear chaotic oscillators. This study showed that by distributed sensing and stimulation seizures can be suppressed, which is consistent with clinical studies [24]. A novel seizure control system was developed that consists of a closed loop feedback control system receiving EEG signals and converting them to dynamical indicator quantities known as the short-term maximum Lyapunov exponent (STL_max), which in return activates a medication delivery to prevent seizure occurrence. The success of the technique was demonstrated on rodents with future plans to extend it to human patients [25]. A similar study incorporates average angular frequency from EEG signals together with STL_max and it was shown that both indicators converge preceding seizures in the chronic limbic epilepsy in rats [26]. As an extension to this study an electrical stimulation seizure control system based on state space regional coupling was developed [27]. STL_max was also used as a new brain mapping method using a Gaussian mixture model to approximate the spatial distribution of STL-max. The results provided important information on the spatial organization of the epileptogenic focus dynamics [28].

Quantitative methods were especially emphasized in visualization methods, some of which involved complex mathematical techniques. A variety of geometric invariants that quantify topological properties of cortical surfaces were introduced on global and local levels of detail in 3D space [29]. Including the time component, through a similar study, a visualization tool was developed to help spatiotemporal analysis on neural models, voltage-sensitive-dye imaging and multi-electrode array experiments. The effectiveness of the tool is shown on a comparison of structure, timing and synchronization properties of neural populations [30]. Another visualization technique was developed to track changes in the neural tissue after the insertion of a prosthetic device. A finite-element method based model was used to modulate parameters of the device with respect to the changing geometry of the neural tissue [31].

5- Conclusions

Significant results on all of the main topics of optimization, system analysis and data mining were presented. Contributions ranging from theoretical and computational neuroscience to detection and prevention of neural abnormalities and to novel techniques of visualization on different scales point to new research directions, while possible interactions between the research areas define new dimensions to be explored in neuroscience. With such a great variety of different approaches, and highly interconnected problems, the conference concluded with the exchange of many ideas and results, which are promising seeds for fruitful collaborations in new research directions. The productivity of this conference has been encouraging and inspirational to many researchers. Future conferences will continue to bring together the prominent researchers, who are on the cutting edge of the growing field of neuroscience.

6- Open Questions And Future Directions

• Is there an underlying neural signal code that governs brain functions? Or is there no code at all?
• How is the brain organized from micro to macro scale? What are the different stages of the transition between these scales?
• What are the evolutionary mechanisms that shaped the wiring in the brain? What are the similarities and differences between wiring structure across different species?
• How are the functions of the brain integrated over time and over different cortical areas in different scales?
• Is it possible to detect a specific thought from EEG recordings?
• Which channels are most relevant to perform data mining? How can such channels be chosen in an automatic way?
• How can the source of evoked potentials be precisely localized?
• Is it possible to detect epileptic seizures with 100 percent accuracy? Is it possible to eliminate seizures by stimulation of the brain.
• Can the brain functions be modeled accurately based on statistical methods only?
• How can the brain function be visualized at the microscopic level? How can visualization tools be integrated at different levels of resolution?
• How can the brain be simulated accurately to produce similar input/output even for basic functions?

7- Acknowledgements

The author and the DIMACS Center acknowledge the support of the National Science Foundation under grant number NSF CCF 05-14703 Workshops Connecting Theoretical Computer Science to Other Fields to Rutgers University.

References

[1] Walter J. Freeman, Sourcing organizing concepts for neocortical dynamic data from many-body physics. University of California, Berkeley

[2] Andreas Ioannides, Probing Brain Function Across Different Spatial and Temporal Scales with Tomographic Analysis of Magnetoencephalographic Signals. Brain Science Institute, RIKEN, Japan

[3] Dmitri Chklovskii, Evolution as the Blind Engineer: Wiring Minimization in the Brain. Cold Spring Harbor Laboratory

[4]Fred Glover, Discrimination and Classification by Mixed Integer Programming. University of Colorado

[5] Michele Samorani, Hyperplane-Based Decision Trees and Their Optimization Università degli Studi di Bologna, Italy

[6] Fang Liang, Computational Evaluation of Mixed Integer Programming Models for Discrimination and Classification. University of Colorado

[7]Onur Seref, Kernel Based Methods Applied to Single Trial Neural Signals. University of Florida

[8] Larry Manevitz, Reading the Mind: fMRI Analysis Via One-Class Machine Learning Techniques. University of Haifa, Israel

[9] Alexander Hartemink, Neural Information Flow Networks in Songbirds. (Duke University

[10] Xue Bai, Tabu Search Enhanced Graphical Models for Classification of High Dimensional. Carnegie Mellon University

[11] Pando Georgiev, Fuzzy Hyperplane Clustering Algorithm and Applications to Sparse Representations. University of Cincinnati

[12] Moongu Jeon , Parallel Image Clustering using Level Set Methods. Gwangju Institute of Science and Technology, Korea

[13] W. Art Chaovalitwongse, Cluster Analysis of Epileptic Brains. Rutgers University

[14] Mingzhou Ding, Statistical Modeling of Neurobiological Data. University of Florida

[15] Mukesh Dhamala, Current Source Density Analysis of Ongoing Neural Activity: Theory and Applications. University of Florida

[16] Linda Hermer-Vazquez, Adaptations of Stationary Analytics Techniques to Understand the NonstationaryNeural Activity Underlying Performance of a Complex Cognitive Task. University of Florida

[17] Anant Hegde, Tracking Spatio-Temporal Changes in ECOG. University of Florida

[18] Jeff Knisley, Neural Networks, Monte Carlo Methods, and Real-world Neurons. East Tennessee University

[19] Maya Maimon, A Simulation Tool Using Discrete Integrate and Fire Neurons: Modeling the Influence of Anatomy on Information Flow in Very Large Simulated Networks. University of Haifa, Israel

[20] Richard E. Frye, Dyslexia: An Example of Natural Variation in Large-Scale Neural Network Organization. University of Florida

[21] Yan Zang, MEG in Dyslexia: A Power and Coherence Study. University of Florida

[22] Wichai Suharitdamrong, Graph Theory-Based Data Mining Techniques to Study Similarity of Epileptic Brain Network. University of Florida

[23] Leonidas D. Iasemidis, Resetting of Brain Dynamics by Epileptic Seizures. Arizona State University

[24] Kostas Tsakalis, A Feedback Control Systems View of Epileptic Seizures. Arizona State University

[25] Kevin M. Kelly, Development of a Rodent Seizure Control System using IntracerebroventricularInjections of Midazolam. Drexel University

[26] Paul R. Carney, Dynamical EEG Properties in the Limbic Epilepsy Rat Model. University of Florida

[27] Sandeep P. Nair, Dynamical State Dependent Electrical Stimulation for Seizure Control in a Chronic Limbic Epilepsy Model. University of Florida

[28] Nadia Mammone, A New Brain Mapping Based on the Visualization and Modeling of the Short TermMaximum Lyapunov Exponent. Reggio Calabria, Italy

[29] Monica K. Hurdal, Shape Analysis for Automated Sulcal Classification and Parcellation of MRI Data. (Florida State University

[30] Kay A. Robbins, Visual Analysis for Comparing Structure, Timing and Synchronization Properties of Neural Populations. Univsersity of Texas at San Antonio

[31] Aparna Gupta, Online Analysis of Device-Tissue Interactions - Modeling Tissue Impedance Spectra. Resselaer Polytechnic Institute