Artificial Intelligence in Biomedical

Engineering

DRAGOS AROTARITEI
[email protected]

University of Medicine and Pharmacy “Grigore T. Popa” Iasi

Department of Medical Bioengineering
Romania

May 2016
Introduction
 Artificial intelligence (AI) is an area of computer science, which has been
developed since the 1950s, specialized in dealing with problems considered
difficult by traditional computer scientists through the use of knowledge and
of probabilities and other kinds of uncertainties.
 AI techniques can be applied to solve complex problems in biomedical
engineering.
 AI subfields
 Neural Networks
 Evolutionary Computation,
 Vision
 Robotics
 Expert Systems
 Speech Processing
 Planning,
  Machine Learning Algorithms can create new medical knowledge
(supervised and unsupervised)
 Natural Language
 Fuzzy systems and approximate reasoning
 Hybrid systems
 Machine Learning algorithms
 UC Irvine Machine Learning Repository
 http://archive.ics.uci.edu/ml/
 Pattern recognition - a branch of machine learning that focuses on the
recognition of patterns
Feature extraction
 Feature extraction is an intermediary step between preprocessing data and the
process (classification, clusterization, rule generation, etc.)
 Two problems arise after data is collected: dimensionality of each item from
data and the number of items. Sometimes both problems are possible to be
solved in a single step.
 The digital information is usually structured in vectors and matrix having
sometimes very large dimension. As consequence, manipulation of these types
of data requires a dimensionality reduction: principal component analysis
(PCA), linear discriminant analysis (LDA), canonical correlation analysis
(CCA), singular value decomposition (SVD), dimensionality reduction via
kernels.
 The second problems refers to selection of features from all the features that
are relevant for our goal.
 independent component analysis (ICA), ISOMAP with extensions, kernel
principal component analysis (kernel PCA), Latent semantic analysis (LSA),
Partial least squares regression (PLS regression), ortogonal transformations
orthogonal transformation is a linear transformation T : V → V on a real inner
product space V, that preserves the inner product), nonlinear dimensionality
reduction (NLDR) - manifold learning algorithms
 The are many overlapping algorithms that map the both objectives.
 The feature extraction is can be completely automated very rarely, the ability
of author plays an important role in identification and selection of possible
relevant feature and the method (new or classic one) that leads to results.
Evolutionary Algorithms
 Subset of evolutionary computation
 The most frequently usage: optimization (single objective and multiobjective)
 Genetic algorithms (based on Darwinian evolution) – the most used in
numerical optimizations
 Differential evolution – Based on vector differences
 Ant colony optimization – suitable for path optimization problems
 Particle swarm optimization
 Genetic programming – suitable for finding a mathematical formula for
relationship among variables based on experimental data
 Memetic algorithms - most MAs can be interpreted as a search strategy in
which a population of optimizing agents cooperate and compete using local
strategies
 Other non-Darwinian algorithms: Baldwinian Evolutionary Algorithms,
Lamarckian Evolutionary Algorithms, Genetic Local Search
GA (Genetic Algorithms)
 Used most frequently for numerical optimization problems
 GA belong to heuristic search that mimics the natural evolution (Darwinian
evolution)
 A large area of applications including Biomedical Engineering, Clustering,
Computational Biology, Image processing (Dense pixel matching).
 Explanations of the power of genetic algorithms is given by Holland's schema
theorem (fundamental theorem of genetic algorithms). Low-order schemata
with above-average fitness increase exponentially in successive generations.
A schema is a template that identifies a subset of strings with similarities at
certain string positions1.
 The chromosome is used to code de information. The usually conversion is
value to binary string of „0‟s and „1‟s. Each value „0‟/‟1‟ is named allele.
 A general schema for GA

A general schema for GA (a practical approach)

 Encode/Decode is shortcut for convert values to chromosome/ convert
chromosome to values.
 The algorithm start with an initial number of solutions (initial population)
represented by chromosomes.
 The key problem in GA is the definition of fitness, that usually means the
function minimization (e.g. mean squared error), fitness = y. The algorithm
remain valid for maximum of function if we replace fitness = 1-y.
 The GA operators are: Selection, Crossover and Mutation.
 Selection
 Roulette wheel selection - This fitness level is used to associate a probability of
f
selection with each individual chromosome i
p N
i

f
i 1
i

 Tournament selection - tournament selection is often implemented in practice due

to its lack of stochastic noise
 Stochastic universal sampling
 Reward-based selection
  Boltzmann selection
 Truncation selection
 Fair coin (a sequence of independent Bernoulli trials with probability 1/2 of
success on each trial), p=rand() in Matlab. If p<0.5, k =„1‟ else k = „0‟. This is
often mentioned as Coin flipping.
 Crossover (mate).
 One-point crossover
 Two-points crossover
 “Cut and splice“ – results in change of the length in children chromosomes
 Uniform Crossover (UX) - uses a fixed mixing ratio between two parent
 Half Uniform Crossover (HUX) - only half of the bits that are different will be
exchanged using number of different bits (Hamming distance) between the
parents. The half of this number is the number of bits is exchanged between
parents to form the childs.
 Other that not mimics the nature (Partially-mapped crossover (PMX), Order
crossover (OX) )
 Mutation is a genetic operator used to maintain genetic diversity of population

 Mutation occurs during evolution according to a probability defined by user.

The usually mutation is made by „1‟  „0‟
 Depending on genome (an organism that have a complete set of DNA, that is
all the genes) type mutation can be:
 Bit string mutation
 Uniform
 Non-uniform
 Gaussian
 Boundary (used for integer and float genes)
 Problems and limitations
 Premature convergence
 Elitism selection can produce a uniform population (diversity is at low level) and
population cannot evolve toward a optimal solution
 For high dimensional problems fitness can be expansive
 GA algorithms do not scale well with complexity
 Convergence to local optima instead of global optima (solution the penalty
method)
 For other specific problem some alternative methods can have a faster speed of
convergence, e.g. integer linear programming.
 Multiobjective problems can create difficulties in defining a fitness function
that can have conflicting objectives – Pareto optimal
 Constraints (inequalities, boundaries of topologic surfaces, of n-dimensional
boundaries) can be included in GA
 Other operators has been proposed to improve GA.
Parameter optimizations2
 One of the most important tasks for AF detection is he parameter set for
classifier that are usually thresholds involved in rules or other parameters like
window length.
 Preprocessing stage (Beat sequence from 8219 MIT AFIB database with
ectopic beats and replacement of them)

2D. Arotaritei,
C. Rotariu, Parameter Optimizations in Detection of Atrial Fibrillation Based on Heart Rate Interval,
BIOMEP 2015 - Conference on Bio-Medical Instrumentation and related Engineering and Physical Sciences, 2015,
Athenes. Greece
 The set of parameters are focused on length of window L because an optimal
window can produce a relevant value of randomness of time series that can be
interpreted as AF
 A set of parameters can be defined as P= [L, p1, p2,..., ph]. Optimization of
some parameters can be conflictual that is if optimization of objective using
one parameter can produce a decrease of optimized value for other
parameters. A tradeoff using optimized curve of values can be used to choose
a correct tradeoff (e.g. Pareto front).
 The selected parameters are3: L – length of window, Pth1 – threshold from
where the segment is considered to be AF (number of AF segments/number of
total segments), RMSt – threshold of Root Mean Square of Successive
Differences, TPRt threshold of Turning Point Ratio, SEt - threshold of
Shannon Entropy and TKt - threshold of Teager-Kaiser energy (six
parameters).
 An vector of HRV descriptors are good for detection of AF if sensitivity Se
and specificity Sp are both enough close to 1.0 for a correct classification.
 Genetic Algorithms proved to be a good choice for optimization problem,
single objective or multi-objective.
 The parameters are set to be limited by lower values [32, 0.1%, 0.01, 0.01%,
0.01, 0.01] and upper values [480, 99.99%, 0.999, 99.99%, 0.999, 0.999].
 The chromosome coded for first value is mapping linearly into integer values
between 32 and 480 (logarithmic mapping is a solution for some types of
problems).
 We start with a population of 50 individuals and after 40 iterations the
solution is given in next figure. The multi-objective function is minimization
of bot functions 1- Se and 1-Sp. The results are Se=90.17 and Sp = 90.12.
Single objective GA, ffitness= 1-min(Se, Sp)
fusion of two objectives (if we have more than two
objectives, the graphic become more difficult).
Artificial Neural Networks (ANN or shortly NN)
 Neural networks are models inspired by biological neural networks. Main
features: neuron transfer function, architecture, neuron connections.
 Main usage: function approximation, classifier, pattern recognition,
prediction.
 ANN proven to be universal approximators (Haykin S.).
 In 1989, Cybenko G. demonstrated, using a Kolmogorov‟s older result that
multilayer feed-forward network with a single hidden layer, which contains
finite number of hidden neurons, is a universal approximator. The
demonstration has been made for sigmoid activation function.
 In the most common sense, the universal approximator can approximate any
non-linear function with a desired precision if its architecture is large enough.
 Other NNs architectures have been proved to be also universal approximators:
radial basis function (RBF) (Park J., and Sandberg I.W.), recurrent neural
networks (RNN) (Schäfer A.M., and Zimmermann H.G.) and Kohonen maps.
 Recurrent Neural Networks (RNN) in various architectures and connection
among neurons are the subject of dynamic nonlinear modelling and prediction
with very good results.
 The most common architecture is feedforward Neural network (multilayer
perceptron) a (axial depth of cut [mm]), f (feed per tooth
p z
[mm/tooth]), and Vc (cutting speed [m/min]).
The outputs layer has six neurons
corresponding to average cutting forces and
maximum measured forces

The output of neuron in the layer k is given by

an activation function that takes into account all
the output of neurons of layer k  1, weighted
and biased by neuron parameters.
 n j k k 1 k 
x     wij x j  bi    ( wT x  b)
k
i  j 1 
 
 ( x)  1/(1  e x b )
Model of nonlinear phenomenon using ANN
 Grinding is an essential procedure in orthopedic surgery.
 High-speed miniature grinding wheels are often used to remove material.
 E.g. Stryker CORE U Drill with wheel, Diamond Round Bur type (round
shape)
wijk (t  1)  wijk (t )  wijk (t )

 Tabular data are available from experiments

 The inputs are linearly scaled in range [-1, +1] in order to prevent saturation
od sigmoidal functions.
Learning algorithm: backpropagation with
clustering data.
Backpropagation – gradient based algorithm
Two stages: feedforward data
backpropagation error
Weights are initially set to random values and
updated in these two stages by w (t  1)  w (t )  w
k
ij
k
ij
k
ij
(t )
R2 - coefficient of determination, proportion of the
variance in the dependent variable that is
predictable from the independent variabl
Conclusions
 Heuristic algorithms are a very useful solution in many biomedical
applications
 Preprocessing experimental data plays an important role in success o chosen
algorithm
 Feature selection are most often depending on the author‟s choice
 If n-dimensional data are for two ore more classes are spatially too close,
some operation (e.g. orthogonalization) are necessary.
 Some GA not truly mimic the nature. Some operations and operators are
introduced based statistical and mathematical consideration in order to find a
acceptable solution for a given problem (minimization and maximization)
 Genetic algorithms start with a random "gene" sets (chromosomes). In the real
world an organism with random genes would not live.
 The initial population is like a set amino acids form a kind of primordial soup
References
 Nikolaos G. Bourbakis, Artificial Intelligence Methods and Applications, World Scientific, 1992.
 Stuart J. Russell and Peter Norvig, Artificial Intelligence A Modern Approach, Pearson, 2009.
 Mariusz Flasinski, Introduction to Artificial Intelligence, Springer, 2016.
 http://infolab.stanford.edu/~ullman/mmds/ch11.pdf
 Zhisong PAN , Zhantao DENG, Yibing WANG, Yanyan ZHANG, Dimensionality reduction via kernel sparse
representation, Front. Comput. Sci., 2014, 8(5): 807–815
 Zbigniew MichalewiczGenetic Algorithms + Data Structures = Evolution Programs, Springer, 1996
 David E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-Wesley
Professional, 1989
 Carlos Coello Coello, Gary B. Lamont, David A. van Veldhuizen, Evolutionary Algorithms for Solving
Multi-Objective Problems, Springer, 2007.
 Mehryar Mohri, Afshin Rostamizadeh, Ameet Talwalkar, Foundations of Machine Learning, The MIT Press,
2012.
 George J. Klir, Bo Yuan, Fuzzy Sets and Fuzzy Logic: Theory and Applications, Prentice Hall, 1995
 Simon Haykin, Neural Networks: A Comprehensive Foundation (2nd Edition), Prentice Hall, 1998.
 https://arxiv.org/ftp/arxiv/papers/1408/1408.0101.pdf
 http://www.lcc.uma.es/~ccottap/papers/handbook03memetic.pdf
 1J. Holland, Adaptation in Natural and Artificial Systems, The MIT Press, 1992.

 S. Dash, K.H. Chon, S. Lu, E.A. Raeder, “Automatic Real Time Detection of Atrial Fibrillation”, Annals of
Biomedical Engineering, vol. 37, issue 9, pp. 1701-1709, 2009.
Conclusions
 A gene is a segment of DNA containing the code used to synthesize a protein.
A chromosome contains hundreds to thousands of genes.
 The number of individuals in population at each generation is fixed GA. Some
researchers proposed GA schemes using variable size of population during
evolution
 The length of chromosome are the same for all individuals. There are some
proposals that use a variable length genotype (the organism itself), SAGA.
 Algorithms from Artificial Intelligence are not a universal solution for
optimized problems - sometime other solution based on non-heuristic models
can be more efficient.
Σας εσταριστώ για την προσοτή σας