Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
  • G06F17/10—Complex mathematical operations
  • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N7/00—Computing arrangements based on specific mathematical models
  • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/08—Speech classification or search
  • G10L15/14—Speech classification or search using statistical models, e.g. Hidden Markov Models [HMMs]
  • G10L15/142—Hidden Markov Models [HMMs]
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
  • G06F2218/12—Classification; Matching

Definitions

• the invention relates in general to modeling and in particular to generating hidden Markov models from state and transition data.
• HMMs Hidden Markov models
• problems that naturally give rise to such data include robot navigation, machine vision, and signal processing, and HMMs are at the core of many state-of-the-art algorithms for addressing these problems.
• many problems of natural language processing involve time-series data that may be modeled with HMMs, including: part-of-speech tagging, topic segmentation, speech recognition, generic entity recognition, and information extraction.
• HMM technology appears in numerous fields including and not limited to voice recognition, handwriting recognition, signal processing and genetic engineering. It is a fundamental tool for uncovering state systems within complex data sets of real world phenomena. However, many techniques for arriving at a HMM representative of such complex data are highly empirical. Thus, there is a need for improved methods to generate a HMM from such data sets, to test and/or change the complex systems in accordance with the HMM.
• This invention arises from studies of mouse sleep stage data, iterating related art techniques originally designed for studying ion channels ("Maximum likelihood estimation of aggregated Markov processes" Proceedings of the Royal Society B, Vol.
• this invention presents a method to arrive at the "best" or most likely graphical model.
• This method is a data processing technique to identify hidden Markov model (HMM) state machines in physical, chemical, biological, physiological, social and economic systems.
• HMM hidden Markov model
• this invention does not choose from a library of pre-determined left-to-right models, or any other library, but determines a new model from each new set of data.
• a state machine is a concept that is used to describe a system that transitions from one state to another state and from there back to the original state or into other states, and so on.
• Dwell time is the time spent in any one state. Dwell times and transitions between states can be observed, but they are often aggregations that cannot be distinguished by limited or indirect observations.
• the observed state machine may include invisible transitions among
• transitions are instantaneous and random; the probability per time unit of a transition at a given time from one state to another ideally depends only on the rate of that transition and the state at that time, and not the history of the system. These transition rates allow otherwise identical states to be distinguished, in that states with different exit transition rates will generally have different dwell time distributions. Observations are made over a period known as an epoch, a frame or a sampling interval, and for each of these a class or aggregated state is assigned. The aggregated states thus can easily be distinguished in histograms of their observed dwell times.
• mice Physiological and biological processes often resemble state machines.
• states identified as rapid eye movement (REM) sleep, slow wave sleep and awake are readily identified in EEG polysomnography studies and, at first glance, a simple 3 state machine emerges with transitions between all states (except you don't see transitions directly from awake to REM sleep). The transitions occur randomly without apparent outside stimulus and so the state machine can be considered a Markov system.
• histograms of the 3 observed state dwell times indicate that there are multiple hidden states for each of the observed states. How to connect these 6 or more hidden states with hidden transitions is not at all clear and in fact the number of possible connected graphical models increases combinatorically with the number of states and transitions.
• the hidden Markov model has states and transitions that are not readily apparent from the data but nevertheless are real components of the system that is represented by the Markov model. By uncovering the hidden Markov model, investigators learn more about the underlying processes and are better able to explain the phenomena of studied physical, chemical, biological, physiological, social and economic systems and craft experiments to measure how variables will affect the systems.
• Markov models allow the observer to make predictions about possible results if the system is activated in different ways. For example, data from a control Markov system may be compared with data from an experimental Markov system to see if the variables between the control and experimental systems generate changes on the system level, i.e., do they create different states and different transitions between states. Comparing control and experimental Markov systems gives more information about not only the gross differences between the control and experimental system but also the way in which those differences are manifested in the operation of the system. In our analysis of very limited mouse sleep data, for example, we discover plausible wild-type mouse sleep cycles, and that the double knock-out mice have dramatic changes in their sleep models, a result that could not be determined by gross observation of single knock-out mice (see Joho).
• the present invention is directed to a data enhancement method for the presentation of data for improved human reading, analyzing and/or interpreting.
• the present invention is directed to the problem to present data in a manner, which improves the readability, the ability to analyze and/or the ability to interpret data, enabling the user to perform his task more efficiently.
• the present invention relates to how cognitive content is conveyed to the reader, analyzer and/or interpreter.
• the present invention overcomes the above-mentioned problems of the prior art and allows investigators to find hidden Markov models by following a set of rules.
• the rules exploit and follow the data in a given data set so that the investigator performs a series of repeated steps that lead to a "best" (e.g., most likely) hidden Markov model at each iteration.
• the best candidate model(s) At the end of each iteration of the rules of the steps, the best candidate model(s) have been stored, and their score (e.g., likelihood) is compared with that of the next best candidate model(s). If the difference in scores is significant, then the additional complexity of the best candidate(s) is justified.
• This invention is based on a combination of statistical probability and Markov model structures, their construction and their modification that is driven by data under examination.
• the invention identifies isomorphic (identical or redundant) models and analyzes only one isomorphic model during an iteration of the steps.
• the rules allow for some variation in their application— at the outset, with the choice of initial model, and along the way, and if problems are encountered.
• the rules are robust, in that the same result is usually obtained by different applications of the rules (e.g., different choices of equally "best” candidates along the way, or different choices of starting model).
• the invention can visualize the found Markov models to the user, wherein the states and transitions of each Markov model are arranged as items (or images) on a screen or a printout or the like: States may be visualized by items or symbols (e.g. a rectangular box), transitions may be visualized by arrows connecting the states and transitions probabilities may be visualized by numbers.
• States may be visualized by items or symbols (e.g. a rectangular box), transitions may be visualized by arrows connecting the states and transitions probabilities may be visualized by numbers.
• This invention provides a tool to, inter alia, characterize and visualize the physiology of various organisms, i.e. the invention allows to determine the physical, chemical, biological, biochemical and/or psychological functions and processes of the corresponding organism.
• the organism can be a 'living' system such as a molecule, a bio-molecule, a cell, an organ or the like.
• the organism can represent any biological process of an organism.
• the invention can be used to determine the effectiveness and/or
• the present method is understood as a computer-implemented invention such as a program (software) installed on a computer.
• the program may process data which represent physical entities such as pharmaceutical products or drugs.
• Figure 1 shows a sleep cycle for wild mice in the dark.
• Figure 2 shows a sleep cycle for wild mice in light.
• Figure 3 shows a sleep cycle for genetically altered mice in the dark.
• Figure 4 shows a sleep cycle for genetically altered mice in light.
• Figure 5 shows an initial model for uncovering a Hidden Markov Model.
• Figure 6 is one way of adding a new transition.
• Figure 7 is another way of adding a new transition.
• Figure 8 is a starting guess that achieves the next best status.
• Figure 9 is an optimized model derived from the starting guess of Figure 8 that achieves the next best status.
• Figure 10 is a starting guess for the optimization that achieves the best status.
• Figure 11 is an optimized model derived from the starting guess of Figure 10 that achieves the best status.
• Figures 12 and 13 show the next two steps of growth.
• ModelGrower.py a Python script that runs in a convenient interface provided by QUB. A copy of the source code for this program is appended to this patent.
• geng.exe allpermg.exe, shortg.exe, and listm.exe in the NAUTY22 directory, and associated extensions/modifications, straightforward to someone skilled in the art, of Brendan McKay's open source software package NAUTY to properly handle color partitions, the original having been obtained online, used for counting and eliminating isomorphic duplicates of graphs underlying the Markov models, and
• a cygwin environment is needed to compile 2) and run 3) on a PC.
• a convenient setup tool for a cygwin environment is available at http://www.cygwin.com.
• Maximum Likelihood Methods have long been used to fit transition rates of hypothetical Hidden Markov Models to observational data. It has been a weakness of these methods that they can only optimize a few parameters in an otherwise rigid model. This invention provides one way to let the data tell us what the model should be, namely what the most likely underlying graph is, without any a priori assumptions.
• This invention solves this problem by letting the data tell us what the most likely addition to the model should be.
• mice The data presented below relies upon two sets of mice.
• One set includes ordinary or so- called wild type mice that have had no genetic alterations.
• the other set include mice that have been genetically altered to remove two genes.
• the latter is set is called double knock out (DKO) mice.
• DKO double knock out mice.
• EEG-based data are spectrally assigned sleep state observations for 24 hours on a 12/12 light/dark cycle for 13 individual wild-type (WT) and 13 individual Kv3.1/Kv3.3 double knock-out (DKO) mice (see files MouseSleepKineticsWT.dwt and MouseSleepKineticsDKO.dwt and the corresponding light and dark selection lists).
• This invention provides a tool to characterize the physiology of these differences in explicit detail, as is already apparent in the models we develop.
• the wild-type mouse sleep cycles for light and dark are very similar, differing mainly in the kinetics of waking states where the sleep cycle is entered and exited (the numbering of the states shows only the order of their addition to the model— the states of each color are indistinguishable aggregated states) as shown in Figs. 1 and 2.
• These represent the HMMs for the most likely candidate after seven steps on the wild type dark data (to WTdark9) and five steps on the wild type light data (to WTlight7), respectively.
• the double knockout mice have radically different sleep models from the wild-type and even from dark to light as shown in Figs. 3 and 4. They represent the HMMs for the most likely candidate after six steps on the DKO dark data (to DKOdark8) and seven steps on the DKO light data (to DK01ight9), respectively.
• Figs. 1-4 show the HMMs found by the inventive method.
• the HMMs are visualized on a screen or printed out on a paper in the way they are shown in the Figs. 1-4.
• the results of the invention show that each of the four sets of data has different HMMs.
• the wild type dark and light are similar to each other. However, the DKO dark and light are different from each other and from the corresponding wild type data for dark and light.
• the invention can readily distinguish between wild type and DKO sleep patterns. While existing methods could not distinguish differences without double knockouts, it seems likely that each single knockout would have caused changes in the sleep cycle that this method would have found, thus elucidating the function of the knocked-out KV3.1/KV3.3 potassium channels.
• the invention is performed by operating one or more programs on a computer. Results are presented on a display or are printed out on a physical entity giving visual indications automatically about conditions prevailing in the system, which is represented by the processed data. In order to follow the steps of the invention, the following notes are provided.
• ModelGrower.py should replace ModelBuilder.py in the PythonScripts directory. Execution Notes:
• Any HMM extracted by the inventive method during the optimization process can be visualized and arranged on a screen or printed out on a paper to inform an operator/user online.
• SUMMARY OF MOUSE SLEEP MODEL GROWTH growth step bestLL nextbestLL deltaLL bestLL nextbestLL deltaLL
• NAUTY program For example, after the first step in the first data set, there are 15
• the first set of data passes through seven steps before reaching a final step where further improvement is not likely.
• Each set of data is processed with and without a hypercube.
• the log likelihood (LL) of each member of each step is provided by the QUB program. Only the best and next best are shown in the above table.
• the data is tested with and without hypercubes of starting values.
• a blank cell under the non-hypercube column indicates the results for the non-hypercube are the same as the results for the hypercube. When results are different, the results are shown in the non-hypercube column.
• the delta LL shows the differences between the best one of the distinct graphs and the next best one for each step of graphs.
• the first two sets of data reach diminishing returns and this is shown by their respective delta LLs reducing to 1.64 and 0.88, respectively.
• the best graph of the last set is selected as the most likely HMM (at the fourth steps, in these cases).
• the observed or aggregated states include REM state 1 assigned the color black and indicated as a square in the figures, Slow Wave State 2 assigned the color red and designated by an oval, and the waking state 3 assigned the color blue and designated by a hexagon. Note that we could have included a transition between REM state 1 and waking state 3, but there are actually no transitions from waking to REM in the data, so we chose not to include this transition in either direction at this stage.
• the invention uses the ModelGrower program to generate possible candidates from an original or aggregated model.
• the NAUTY program operates on the candidates to identify and exclude isomorphic models.
• the QUB program then operates on the remaining non-isomorphic candidates to identify the candidate that most closely conforms to the data by optimizing rates of that candidate (e.g., maximizing the likelihood that the data came from the model with those rates).
• the optimized candidate of the first stage is the starting candidate for the next stage where ModelGrower, NAUTY and QUB operate again.
• the invention terminates at an end point defined by the user, preferably an end point with threshold determined by diminishing delta LL.
• the ModelGrower program performs the process of growing the basic model to candidate models that are representative of all possible models with one more transition.
• ModelGrower program starts with the basic observed model of Fig. 5 and grows it by splitting states or connecting original unconnected states.
• NAUTY reduces the number of candidates to 15 non-isomorphic candidates with one additional transition.
• QUB examines the 15 candidates and ModelGrower selects the one candidate that most closely conforms to the data.
• the best candidate of stage 1 becomes the new starting point for stage 2 and it is examined for hidden states and hidden transitions.
• the program ModelGrower grows the candidate in all possible ways by first splitting existing (aggregated) states into two states of the same color. NAUTY removes isomorphs.
• ModelGrower splits each state into two states where one split state is the same color as the original state and the other split state is a different color.
• the number of colors corresponds to the initial number of observed, aggregated states.
• NAUTY operates on those states to remove isomorphs.
• ModelGrower connects all unconnected states and NAUTY operates once more.
• QUB evaluates the 15 candidates by optimizing them, and ModelGrower identifies the best one of the candidates.
• the optimized candidate then becomes the starting candidate for stage 2 where the candidate is again grown by ModelGrower into more candidates, those candidates are examined for isomorphs by NAUTY to reduce the large number of possible combinations to 44 and those 44 candidates are optimized by QUB.
• the process is repeated seven times until one reaches an end point. One may set the end point at any suitable threshold. For the wild type dark data the end point was selected where the next delta LL was 5.16. That indicates the improvement in data for the model is minor.
• Figs. 6 and 7 exemplify two ways of adding a new transition.
• Fig. 6 shows how a prior state 2 colored red (oval) may be separated into another state 4, of the same color (red, oval) and into a new state 2 of the same color (red, oval).
• Fig. 7 shows how a transition is added between prior state 3, blue (hexagon) and state 1, black (square).
• Figures 6 and 7 are just two of the 15 possible non-isomorphic evolutions of the primary or original aggregated model of Fig. 5. These starting rates are exemplary and other starting rates may be used. Note that all other transition rates have been retained as starting values for the optimization by QUB of these working models.
• 15 non-isomorphic graphs are built in this way as starting guesses for optimization by QUB to find the next biggest model.
• Those 15 models are optimized by QUB and ModelGrower selects the best one of the 15 initial models and that model becomes the new model for the next iteration of the invention.
• the best or optimized model of in each step is used to generate the models of the next step. Those models have their isomorphs removed by NAUTY and the non-isomorphs are optimized with QUB so ModelGrower may select the best model for the next step. The foregoing process is repeated until there is little or no improvement.
• the invention operated seven times for wild type dark, five times for the wild type light, six times for the DKO dark and seven times for the DKO light.
• the final, optimized models are shown in Figs. 1-4. Any content shown in Figs. 1-13 is displayed to a user/operator on a screen.
• the displayed HMMs present relations within the processed data: the various states are visualized by symbols (squares, circles, ...) with integer numbers, transitions between these states are visualized by arrows and transition probabilities are visualized by floating point numbers beside the arrows.
• the invention is concerned with a computer-implemented system for processing data input to the system by a user.
• the system processes the data and generates HMMs.
• the HMMs consist of numbers comprising two or more digits, which identify states and transition probabilities.
• the user can use the information/cognitive content from these HMMs to adapt his organism under research, i.e.
• the user can respond in terms of modifying the organism by changing a single parameter characterizing the organism, by extracting new observation data representing the modified organism and by sending the new data to the inventive system for reprocessing.
• the reply data includes code identifying the modified organism as well as a word (or part of a word) composed of the letters representing the modification identified by the code digits.
• the present invention also relates to a computer readable computer program product with computer executable instructions prompting the computer to execute the aforementioned methods or processes.
• the computer program product can be a CD, DVD, HDD, USB-stick, memory card (CF, SD, MicroSD, MiniSD, SDHC, ...) or the like with a suitable software program recorded on it.
• the attached Appendices provide detailed steps for operating their respective programs.
• the QUB and NAUTY programs are available for use with the invention and they are hereby incorporated by reference. Data and other disclosure in the references discussed above are also incorporated by references.
• the invention uncovers HMMs by assuming that simplistic observed data includes one or more hidden states or hidden transitions between states.
• the invention may be used to generate HMM from complex data, especially data representative of biological processes.
• the invention provides valuable tools and processes to investigate the structure and operation of such processes. There are numerous applications.
• One example is ion channel communication. Physiologists believe that ion channels in cells control intercellular and intracellular
• HMM for ion channel operation.
• the structure of the state machine it may be possible to treat a disease by using one or more medicines, electrical potential or currents or physical perturbations to alter a state or a transition between states.
• a disease is characterized by an over abundance of an immune response and the body produces an excess of cytokines and that over production is harmful.
• the HMM is may be possible to uncover a key state or key transition that may be manipulated by chemical, electrical, mechanical or other means to alter the state or transition and thereby mute the response.
• Another example is the opposite case with HIV where the body is deficient in its immune response.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• Data Mining & Analysis (AREA)
• Theoretical Computer Science (AREA)
• Mathematical Analysis (AREA)
• Computational Mathematics (AREA)
• Mathematical Optimization (AREA)
• Mathematical Physics (AREA)
• Probability & Statistics with Applications (AREA)
• Pure & Applied Mathematics (AREA)
• Software Systems (AREA)
• General Engineering & Computer Science (AREA)
• Algebra (AREA)
• Databases & Information Systems (AREA)
• Human Computer Interaction (AREA)
• Evolutionary Biology (AREA)
• Bioinformatics & Computational Biology (AREA)
• Bioinformatics & Cheminformatics (AREA)
• Life Sciences & Earth Sciences (AREA)
• Computational Linguistics (AREA)
• Health & Medical Sciences (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Operations Research (AREA)
• Acoustics & Sound (AREA)
• Multimedia (AREA)
• Artificial Intelligence (AREA)
• Evolutionary Computation (AREA)
• Computing Systems (AREA)
• Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
• Management, Administration, Business Operations System, And Electronic Commerce (AREA)
• User Interface Of Digital Computer (AREA)
• Complex Calculations (AREA)

Applications Claiming Priority (1)

Publications (2)

Family

ID=46831013

Family Applications (1)

Country Status (6)

Families Citing this family (1)

Family Cites Families (5)

• 2011
  • 2011-03-14 WO PCT/US2011/028302 patent/WO2012125146A1/en active Application Filing
  • 2011-03-14 SG SG2013069059A patent/SG193450A1/en unknown
  • 2011-03-14 AU AU2011362611A patent/AU2011362611B2/en active Active
  • 2011-03-14 JP JP2013557997A patent/JP5898704B2/ja active Active
  • 2011-03-14 EP EP20110861132 patent/EP2686796A4/de not_active Ceased
  • 2011-03-14 CA CA2830159A patent/CA2830159C/en active Active

Also Published As

Similar Documents

Legal Events